Tree trunk instance segmentation is crucial for under-canopy unmanned aerial vehicles(UAVs)to autonomously extract standing tree stem attributes.Using cameras as sensors makes these UAVs compact and lightweight,facili...Tree trunk instance segmentation is crucial for under-canopy unmanned aerial vehicles(UAVs)to autonomously extract standing tree stem attributes.Using cameras as sensors makes these UAVs compact and lightweight,facilitating safe and flexible navigation in dense forests.However,their limited onboard computational power makes real-time,image-based tree trunk segmentation challenging,emphasizing the urgent need for lightweight and efficient segmentation models.In this study,we present RT-Trunk,a model specifically designed for real-time tree trunk instance segmentation in complex forest environments.To ensure real-time performance,we selected SparseInst as the base framework.We incorporated ConvNeXt-T as the backbone to enhance feature extraction for tree trunks,thereby improving segmentation accuracy.We further integrate the lightweight convolutional block attention module(CBAM),enabling the model to focus on tree trunk features while suppressing irrelevant information,which leads to additional gains in segmentation accuracy.To enable RT-Trunk to operate effectively under diverse complex forest environments,we constructed a comprehensive dataset for training and testing by combining self-collected data with multiple public datasets covering different locations,seasons,weather conditions,tree species,and levels of forest clutter.Com-pared with the other tree trunk segmentation methods,the RT-Trunk method achieved an average precision of 91.4%and the fastest inference speed of 32.9 frames per second.Overall,the proposed RT-Trunk provides superior trunk segmentation performance that balances speed and accu-racy,making it a promising solution for supporting under-canopy UAVs in the autonomous extraction of standing tree stem attributes.The code for this work is available at https://github.com/NEFU CVRG/RT Trunk.展开更多
Semantic segmentation in street scenes is a crucial technology for autonomous driving to analyze the surrounding environment.In street scenes,issues such as high image resolution caused by a large viewpoints and diffe...Semantic segmentation in street scenes is a crucial technology for autonomous driving to analyze the surrounding environment.In street scenes,issues such as high image resolution caused by a large viewpoints and differences in object scales lead to a decline in real-time performance and difficulties in multi-scale feature extraction.To address this,we propose a bilateral-branch real-time semantic segmentationmethod based on semantic information distillation(BSDNet)for street scene images.The BSDNet consists of a Feature Conversion Convolutional Block(FCB),a Semantic Information Distillation Module(SIDM),and a Deep Aggregation Atrous Convolution Pyramid Pooling(DASP).FCB reduces the semantic gap between the backbone and the semantic branch.SIDM extracts high-quality semantic information fromthe Transformer branch to reduce computational costs.DASP aggregates information lost in atrous convolutions,effectively capturingmulti-scale objects.Extensive experiments conducted on Cityscapes,CamVid,and ADE20K,achieving an accuracy of 81.7% Mean Intersection over Union(mIoU)at 70.6 Frames Per Second(FPS)on Cityscapes,demonstrate that our method achieves a better balance between accuracy and inference speed.展开更多
Real-time semantic segmentation tasks place stringent demands on network inference speed,often requiring a reduction in network depth to decrease computational load.However,shallow networks tend to exhibit degradation...Real-time semantic segmentation tasks place stringent demands on network inference speed,often requiring a reduction in network depth to decrease computational load.However,shallow networks tend to exhibit degradation in feature extraction completeness and inference accuracy.Therefore,balancing high performance with real-time requirements has become a critical issue in the study of real-time semantic segmentation.To address these challenges,this paper proposes a lightweight bilateral dual-residual network.By introducing a novel residual structure combined with feature extraction and fusion modules,the proposed network significantly enhances representational capacity while reducing computational costs.Specifically,an improved compound residual structure is designed to optimize the efficiency of information propagation and feature extraction.Furthermore,the proposed feature extraction and fusion module enables the network to better capture multi-scale information in images,improving the ability to detect both detailed and global semantic features.Experimental results on the publicly available Cityscapes dataset demonstrate that the proposed lightweight dual-branch network achieves outstanding performance while maintaining low computational complexity.In particular,the network achieved a mean Intersection over Union(mIoU)of 78.4%on the Cityscapes validation set,surpassing many existing semantic segmentation models.Additionally,in terms of inference speed,the network reached 74.5 frames per second when tested on an NVIDIA GeForce RTX 3090 GPU,significantly improving real-time performance.展开更多
High-resolution remote sensing images(HRSIs)are now an essential data source for gathering surface information due to advancements in remote sensing data capture technologies.However,their significant scale changes an...High-resolution remote sensing images(HRSIs)are now an essential data source for gathering surface information due to advancements in remote sensing data capture technologies.However,their significant scale changes and wealth of spatial details pose challenges for semantic segmentation.While convolutional neural networks(CNNs)excel at capturing local features,they are limited in modeling long-range dependencies.Conversely,transformers utilize multihead self-attention to integrate global context effectively,but this approach often incurs a high computational cost.This paper proposes a global-local multiscale context network(GLMCNet)to extract both global and local multiscale contextual information from HRSIs.A detail-enhanced filtering module(DEFM)is proposed at the end of the encoder to refine the encoder outputs further,thereby enhancing the key details extracted by the encoder and effectively suppressing redundant information.In addition,a global-local multiscale transformer block(GLMTB)is proposed in the decoding stage to enable the modeling of rich multiscale global and local information.We also design a stair fusion mechanism to transmit deep semantic information from deep to shallow layers progressively.Finally,we propose the semantic awareness enhancement module(SAEM),which further enhances the representation of multiscale semantic features through spatial attention and covariance channel attention.Extensive ablation analyses and comparative experiments were conducted to evaluate the performance of the proposed method.Specifically,our method achieved a mean Intersection over Union(mIoU)of 86.89%on the ISPRS Potsdam dataset and 84.34%on the ISPRS Vaihingen dataset,outperforming existing models such as ABCNet and BANet.展开更多
Advanced traffic monitoring systems encounter substantial challenges in vehicle detection and classification due to the limitations of conventional methods,which often demand extensive computational resources and stru...Advanced traffic monitoring systems encounter substantial challenges in vehicle detection and classification due to the limitations of conventional methods,which often demand extensive computational resources and struggle with diverse data acquisition techniques.This research presents a novel approach for vehicle classification and recognition in aerial image sequences,integrating multiple advanced techniques to enhance detection accuracy.The proposed model begins with preprocessing using Multiscale Retinex(MSR)to enhance image quality,followed by Expectation-Maximization(EM)Segmentation for precise foreground object identification.Vehicle detection is performed using the state-of-the-art YOLOv10 framework,while feature extraction incorporates Maximally Stable Extremal Regions(MSER),Dense Scale-Invariant Feature Transform(Dense SIFT),and Zernike Moments Features to capture distinct object characteristics.Feature optimization is further refined through a Hybrid Swarm-based Optimization algorithm,ensuring optimal feature selection for improved classification performance.The final classification is conducted using a Vision Transformer,leveraging its robust learning capabilities for enhanced accuracy.Experimental evaluations on benchmark datasets,including UAVDT and the Unmanned Aerial Vehicle Intruder Dataset(UAVID),demonstrate the superiority of the proposed approach,achieving an accuracy of 94.40%on UAVDT and 93.57%on UAVID.The results highlight the efficacy of the model in significantly enhancing vehicle detection and classification in aerial imagery,outperforming existing methodologies and offering a statistically validated improvement for intelligent traffic monitoring systems compared to existing approaches.展开更多
This systematic review aims to comprehensively examine and compare deep learning methods for brain tumor segmentation and classification using MRI and other imaging modalities,focusing on recent trends from 2022 to 20...This systematic review aims to comprehensively examine and compare deep learning methods for brain tumor segmentation and classification using MRI and other imaging modalities,focusing on recent trends from 2022 to 2025.The primary objective is to evaluate methodological advancements,model performance,dataset usage,and existing challenges in developing clinically robust AI systems.We included peer-reviewed journal articles and highimpact conference papers published between 2022 and 2025,written in English,that proposed or evaluated deep learning methods for brain tumor segmentation and/or classification.Excluded were non-open-access publications,books,and non-English articles.A structured search was conducted across Scopus,Google Scholar,Wiley,and Taylor&Francis,with the last search performed in August 2025.Risk of bias was not formally quantified but considered during full-text screening based on dataset diversity,validation methods,and availability of performance metrics.We used narrative synthesis and tabular benchmarking to compare performance metrics(e.g.,accuracy,Dice score)across model types(CNN,Transformer,Hybrid),imaging modalities,and datasets.A total of 49 studies were included(43 journal articles and 6 conference papers).These studies spanned over 9 public datasets(e.g.,BraTS,Figshare,REMBRANDT,MOLAB)and utilized a range of imaging modalities,predominantly MRI.Hybrid models,especially ResViT and UNetFormer,consistently achieved high performance,with classification accuracy exceeding 98%and segmentation Dice scores above 0.90 across multiple studies.Transformers and hybrid architectures showed increasing adoption post2023.Many studies lacked external validation and were evaluated only on a few benchmark datasets,raising concerns about generalizability and dataset bias.Few studies addressed clinical interpretability or uncertainty quantification.Despite promising results,particularly for hybrid deep learning models,widespread clinical adoption remains limited due to lack of validation,interpretability concerns,and real-world deployment barriers.展开更多
Brain tumors require precise segmentation for diagnosis and treatment plans due to their complex morphology and heterogeneous characteristics.While MRI-based automatic brain tumor segmentation technology reduces the b...Brain tumors require precise segmentation for diagnosis and treatment plans due to their complex morphology and heterogeneous characteristics.While MRI-based automatic brain tumor segmentation technology reduces the burden on medical staff and provides quantitative information,existing methodologies and recent models still struggle to accurately capture and classify the fine boundaries and diverse morphologies of tumors.In order to address these challenges and maximize the performance of brain tumor segmentation,this research introduces a novel SwinUNETR-based model by integrating a new decoder block,the Hierarchical Channel-wise Attention Decoder(HCAD),into a powerful SwinUNETR encoder.The HCAD decoder block utilizes hierarchical features and channelspecific attention mechanisms to further fuse information at different scales transmitted from the encoder and preserve spatial details throughout the reconstruction phase.Rigorous evaluations on the recent BraTS GLI datasets demonstrate that the proposed SwinHCAD model achieved superior and improved segmentation accuracy on both the Dice score and HD95 metrics across all tumor subregions(WT,TC,and ET)compared to baseline models.In particular,the rationale and contribution of the model design were clarified through ablation studies to verify the effectiveness of the proposed HCAD decoder block.The results of this study are expected to greatly contribute to enhancing the efficiency of clinical diagnosis and treatment planning by increasing the precision of automated brain tumor segmentation.展开更多
Modern manufacturing processes have become more reliant on automation because of the accelerated transition from Industry 3.0 to Industry 4.0.Manual inspection of products on assembly lines remains inefficient,prone t...Modern manufacturing processes have become more reliant on automation because of the accelerated transition from Industry 3.0 to Industry 4.0.Manual inspection of products on assembly lines remains inefficient,prone to errors and lacks consistency,emphasizing the need for a reliable and automated inspection system.Leveraging both object detection and image segmentation approaches,this research proposes a vision-based solution for the detection of various kinds of tools in the toolkit using deep learning(DL)models.Two Intel RealSense D455f depth cameras were arranged in a top down configuration to capture both RGB and depth images of the toolkits.After applying multiple constraints and enhancing them through preprocessing and augmentation,a dataset consisting of 3300 annotated RGB-D photos was generated.Several DL models were selected through a comprehensive assessment of mean Average Precision(mAP),precision-recall equilibrium,inference latency(target≥30 FPS),and computational burden,resulting in a preference for YOLO and Region-based Convolutional Neural Networks(R-CNN)variants over ViT-based models due to the latter’s increased latency and resource requirements.YOLOV5,YOLOV8,YOLOV11,Faster R-CNN,and Mask R-CNN were trained on the annotated dataset and evaluated using key performance metrics(Recall,Accuracy,F1-score,and Precision).YOLOV11 demonstrated balanced excellence with 93.0%precision,89.9%recall,and a 90.6%F1-score in object detection,as well as 96.9%precision,95.3%recall,and a 96.5%F1-score in instance segmentation with an average inference time of 25 ms per frame(≈40 FPS),demonstrating real-time performance.Leveraging these results,a YOLOV11-based windows application was successfully deployed in a real-time assembly line environment,where it accurately processed live video streams to detect and segment tools within toolkits,demonstrating its practical effectiveness in industrial automation.The application is capable of precisely measuring socket dimensions by utilising edge detection techniques on YOLOv11 segmentation masks,in addition to detection and segmentation.This makes it possible to do specification-level quality control right on the assembly line,which improves the ability to examine things in real time.The implementation is a big step forward for intelligent manufacturing in the Industry 4.0 paradigm.It provides a scalable,efficient,and accurate way to do automated inspection and dimensional verification activities.展开更多
The application of unmanned driving in the Internet of Things is one of the concrete manifestations of the application of artificial intelligence technology.Image semantic segmentation can help the unmanned driving sy...The application of unmanned driving in the Internet of Things is one of the concrete manifestations of the application of artificial intelligence technology.Image semantic segmentation can help the unmanned driving system by achieving road accessibility analysis.Semantic segmentation is also a challenging technology for image understanding and scene parsing.We focused on the challenging task of real-time semantic segmentation in this paper.In this paper,we proposed a novel fast architecture for real-time semantic segmentation named DuFNet.Starting from the existing work of Bilateral Segmentation Network(BiSeNet),DuFNet proposes a novel Semantic Information Flow(SIF)structure for context information and a novel Fringe Information Flow(FIF)structure for spatial information.We also proposed two kinds of SIF with cascaded and paralleled structures,respectively.The SIF encodes the input stage by stage in the ResNet18 backbone and provides context information for the feature fusionmodule.Features from previous stages usually contain rich low-level details but high-level semantics for later stages.Themultiple convolutions embed in Parallel SIF aggregate the corresponding features among different stages and generate a powerful global context representation with less computational cost.The FIF consists of a pooling layer and an upsampling operator followed by projection convolution layer.The concise component provides more spatial details for the network.Compared with BiSeNet,our work achieved faster speed and comparable performance with 72.34%mIoU accuracy and 78 FPS on Cityscapes Dataset based on the ResNet18 backbone.展开更多
Autonomous driving technology has made a lot of outstanding achievements with deep learning,and the vehicle detection and classification algorithm has become one of the critical technologies of autonomous driving syst...Autonomous driving technology has made a lot of outstanding achievements with deep learning,and the vehicle detection and classification algorithm has become one of the critical technologies of autonomous driving systems.The vehicle instance segmentation can perform instance-level semantic parsing of vehicle information,which is more accurate and reliable than object detection.However,the existing instance segmentation algorithms still have the problems of poor mask prediction accuracy and low detection speed.Therefore,this paper proposes an advanced real-time instance segmentation model named FIR-YOLACT,which fuses the ICIoU(Improved Complete Intersection over Union)and Res2Net for the YOLACT algorithm.Specifically,the ICIoU function can effectively solve the degradation problem of the original CIoU loss function,and improve the training convergence speed and detection accuracy.The Res2Net module fused with the ECA(Efficient Channel Attention)Net is added to the model’s backbone network,which improves the multi-scale detection capability and mask prediction accuracy.Furthermore,the Cluster NMS(Non-Maximum Suppression)algorithm is introduced in the model’s bounding box regression to enhance the performance of detecting similarly occluded objects.The experimental results demonstrate the superiority of FIR-YOLACT to the based methods and the effectiveness of all components.The processing speed reaches 28 FPS,which meets the demands of real-time vehicle instance segmentation.展开更多
The real-time detection and instance segmentation of strawberries constitute fundamental components in the development of strawberry harvesting robots.Real-time identification of strawberries in an unstructured envi-r...The real-time detection and instance segmentation of strawberries constitute fundamental components in the development of strawberry harvesting robots.Real-time identification of strawberries in an unstructured envi-ronment is a challenging task.Current instance segmentation algorithms for strawberries suffer from issues such as poor real-time performance and low accuracy.To this end,the present study proposes an Efficient YOLACT(E-YOLACT)algorithm for strawberry detection and segmentation based on the YOLACT framework.The key enhancements of the E-YOLACT encompass the development of a lightweight attention mechanism,pyramid squeeze shuffle attention(PSSA),for efficient feature extraction.Additionally,an attention-guided context-feature pyramid network(AC-FPN)is employed instead of FPN to optimize the architecture’s performance.Furthermore,a feature-enhanced model(FEM)is introduced to enhance the prediction head’s capabilities,while efficient fast non-maximum suppression(EF-NMS)is devised to improve non-maximum suppression.The experimental results demonstrate that the E-YOLACT achieves a Box-mAP and Mask-mAP of 77.9 and 76.6,respectively,on the custom dataset.Moreover,it exhibits an impressive category accuracy of 93.5%.Notably,the E-YOLACT also demonstrates a remarkable real-time detection capability with a speed of 34.8 FPS.The method proposed in this article presents an efficient approach for the vision system of a strawberry-picking robot.展开更多
The aircraft system has recently gained its reputation as a reliable and efficient tool for sensing and parsing aerial scenes.However,accurate and fast semantic segmentation of highresolution aerial images for remote ...The aircraft system has recently gained its reputation as a reliable and efficient tool for sensing and parsing aerial scenes.However,accurate and fast semantic segmentation of highresolution aerial images for remote sensing applications is still facing three challenges:the requirements for limited processing resources and low-latency operations based on aerial platforms,the balance between high accuracy and real-time efficiency for model performance,and the confusing objects with large intra-class variations and small inter-class differences in high-resolution aerial images.To address these issues,a lightweight and dual-path deep convolutional architecture,namely Aerial Bilateral Segmentation Network(Aerial-Bi Se Net),is proposed to perform realtime segmentation on high-resolution aerial images with favorable accuracy.Specifically,inspired by the receptive field concept in human visual systems,Receptive Field Module(RFM)is proposed to encode rich multi-scale contextual information.Based on channel attention mechanism,two novel modules,called Feature Attention Module(FAM)and Channel Attention based Feature Fusion Module(CAFFM)respectively,are proposed to refine and combine features effectively to boost the model performance.Aerial-Bi Se Net is evaluated on the Potsdam and Vaihingen datasets,where leading performance is reported compared with other state-of-the-art models,in terms of both accuracy and efficiency.展开更多
As the field of autonomous driving evolves, real-time semantic segmentation has become a crucial part of computer vision tasks. However, most existing methods use lightweight convolution to reduce the computational ef...As the field of autonomous driving evolves, real-time semantic segmentation has become a crucial part of computer vision tasks. However, most existing methods use lightweight convolution to reduce the computational effort, resulting in lower accuracy. To address this problem, we construct TBANet, a network with an encoder-decoder structure for efficient feature extraction. In the encoder part, the TBA module is designed to extract details and the ETBA module is used to learn semantic representations in a high-dimensional space. In the decoder part, we design a combination of multiple upsampling methods to aggregate features with less computational overhead. We validate the efficiency of TBANet on the Cityscapes dataset. It achieves 75.1% mean Intersection over Union(mIoU) with only 2.07 million parameters and can reach 90.3 Frames Per Second(FPS).展开更多
This paper concerns the problem of object segmentation in real-time for picking system. A region proposal method inspired by human glance based on the convolutional neural network is proposed to select promising regio...This paper concerns the problem of object segmentation in real-time for picking system. A region proposal method inspired by human glance based on the convolutional neural network is proposed to select promising regions, allowing more processing is reserved only for these regions. The speed of object segmentation is significantly improved by the region proposal method.By the combination of the region proposal method based on the convolutional neural network and superpixel method, the category and location information can be used to segment objects and image redundancy is significantly reduced. The processing time is reduced considerably by this to achieve the real time. Experiments show that the proposed method can segment the interested target object in real time on an ordinary laptop.展开更多
The interpretability of deep learning models has emerged as a compelling area in artificial intelligence research.The safety criteria for medical imaging are highly stringent,and models are required for an explanation...The interpretability of deep learning models has emerged as a compelling area in artificial intelligence research.The safety criteria for medical imaging are highly stringent,and models are required for an explanation.However,existing convolutional neural network solutions for left ventricular segmentation are viewed in terms of inputs and outputs.Thus,the interpretability of CNNs has come into the spotlight.Since medical imaging data are limited,many methods to fine-tune medical imaging models that are popular in transfer models have been built using massive public Image Net datasets by the transfer learning method.Unfortunately,this generates many unreliable parameters and makes it difficult to generate plausible explanations from these models.In this study,we trained from scratch rather than relying on transfer learning,creating a novel interpretable approach for autonomously segmenting the left ventricle with a cardiac MRI.Our enhanced GPU training system implemented interpretable global average pooling for graphics using deep learning.The deep learning tasks were simplified.Simplification included data management,neural network architecture,and training.Our system monitored and analyzed the gradient changes of different layers with dynamic visualizations in real-time and selected the optimal deployment model.Our results demonstrated that the proposed method was feasible and efficient:the Dice coefficient reached 94.48%,and the accuracy reached 99.7%.It was found that no current transfer learning models could perform comparably to the ImageNet transfer learning architectures.This model is lightweight and more convenient to deploy on mobile devices than transfer learning models.展开更多
While traditional Convolutional Neural Network(CNN)-based semantic segmentation methods have proven effective,they often encounter significant computational challenges due to the requirement for dense pixel-level pred...While traditional Convolutional Neural Network(CNN)-based semantic segmentation methods have proven effective,they often encounter significant computational challenges due to the requirement for dense pixel-level predictions,which complicates real-time implementation.To address this,we introduce an advanced real-time semantic segmentation strategy specifically designed for autonomous driving,utilizing the capabilities of Visual Transformers.By leveraging the self-attention mechanism inherent in Visual Transformers,our method enhances global contextual awareness,refining the representation of each pixel in relation to the overall scene.This enhancement is critical for quickly and accurately interpreting the complex elements within driving sce-narios—a fundamental need for autonomous vehicles.Our experiments conducted on the DriveSeg autonomous driving dataset indicate that our model surpasses traditional segmentation methods,achieving a significant 4.5%improvement in Mean Intersection over Union(mIoU)while maintaining real-time responsiveness.This paper not only underscores the potential for optimized semantic segmentation but also establishes a promising direction for real-time processing in autonomous navigation systems.Future work will focus on integrating this technique with other perception modules in autonomous driving to further improve the robustness and efficiency of self-driving perception frameworks,thereby opening new pathways for research and practical applications in scenarios requiring rapid and precise decision-making capabilities.Further experimentation and adaptation of this model could lead to broader implications for the fields of machine learning and computer vision,particularly in enhancing the interaction between automated systems and their dynamic environments.展开更多
Background Generally, it is difficult to obtain accurate pose and depth for a non-rigid moving object from a single RGB camera to create augmented reality (AR). In this study, we build an augmented reality system from...Background Generally, it is difficult to obtain accurate pose and depth for a non-rigid moving object from a single RGB camera to create augmented reality (AR). In this study, we build an augmented reality system from a single RGB camera for a non-rigid moving human by accurately computing pose and depth, for which two key tasks are segmentation and monocular Simultaneous Localization and Mapping (SLAM). Most existing monocular SLAM systems are designed for static scenes, while in this AR system, the human body is always moving and non-rigid. Methods In order to make the SLAM system suitable for a moving human, we first segment the rigid part of the human in each frame. A segmented moving body part can be regarded as a static object, and the relative motions between each moving body part and the camera can be considered the motion of the camera. Typical SLAM systems designed for static scenes can then be applied. In the segmentation step of this AR system, we first employ the proposed BowtieNet, which adds the atrous spatial pyramid pooling (ASPP) of DeepLab between the encoder and decoder of SegNet to segment the human in the original frame, and then we use color information to extract the face from the segmented human area. Results Based on the human segmentation results and a monocular SLAM, this system can change the video background and add a virtual object to humans. Conclusions The experiments on the human image segmentation datasets show that BowtieNet obtains state-of-the-art human image segmentation performance and enough speed for real-time segmentation. The experiments on videos show that the proposed AR system can robustly add a virtual object to humans and can accurately change the video background.展开更多
Instance segmentation plays an important role in image processing.The Deep Snake algorithm based on contour iteration deforms an initial bounding box to an instance contour end-to-end,which can improve the performance...Instance segmentation plays an important role in image processing.The Deep Snake algorithm based on contour iteration deforms an initial bounding box to an instance contour end-to-end,which can improve the performance of instance segmentation,but has defects such as slow segmentation speed and sub-optimal initial contour.To solve these problems,a real-time instance segmentation algorithm based on contour learning was proposed.Firstly,ShuffleNet V2 was used as backbone network,and the receptive field of the model was expanded by using a 5×5 convolution kernel.Secondly,a lightweight up-sampling module,multi-stage aggregation(MSA),performs residual fusion of multi-layer features,which not only improves segmentation speed,but also extracts effective features more comprehensively.Thirdly,a contour initialization method for network learning was designed,and a global contour feature aggregation mechanism was used to return a coarse contour,which solves the problem of excessive error between manually initialized contour and real contour.Finally,the Snake deformation module was used to iteratively optimize the coarse contour to obtain the final instance contour.The experimental results showed that the proposed method improved the instance segmentation accuracy on semantic boundaries dataset(SBD),Cityscapes and Kins datasets,and the average precision reached 55.8 on the SBD;Compared with Deep Snake,the model parameters were reduced by 87.2%,calculation amount was reduced by 78.3%,and segmentation speed reached 39.8 frame·s^(−1) when instance segmentation was performed on an image with a size of 512×512 pixels on a 2080Ti GPU.The proposed method can reduce resource consumption,realize instance segmentation tasks quickly and accurately,and therefore is more suitable for embedded platforms with limited resources.展开更多
Semantic segmentation of eye images is a complex task with important applications in human–computer interaction,cognitive science,and neuroscience.Achieving real-time,accurate,and robust segmentation algorithms is cr...Semantic segmentation of eye images is a complex task with important applications in human–computer interaction,cognitive science,and neuroscience.Achieving real-time,accurate,and robust segmentation algorithms is crucial for computationally limited portable devices such as augmented reality and virtual reality.With the rapid advancements in deep learning,many network models have been developed specifically for eye image segmentation.Some methods divide the segmentation process into multiple stages to achieve model parameter miniaturization while enhancing output through post processing techniques to improve segmentation accuracy.These approaches significantly increase the inference time.Other networks adopt more complex encoding and decoding modules to achieve end-to-end output,which requires substantial computation.Therefore,balancing the model’s size,accuracy,and computational complexity is essential.To address these challenges,we propose a lightweight asymmetric UNet architecture and a projection loss function.We utilize ResNet-3 layer blocks to enhance feature extraction efficiency in the encoding stage.In the decoding stage,we employ regular convolutions and skip connections to upscale the feature maps from the latent space to the original image size,balancing the model size and segmentation accuracy.In addition,we leverage the geometric features of the eye region and design a projection loss function to further improve the segmentation accuracy without adding any additional inference computational cost.We validate our approach on the OpenEDS2019 dataset for virtual reality and achieve state-of-the-art performance with 95.33%mean intersection over union(mIoU).Our model has only 0.63M parameters and 350 FPS,which are 68%and 200%of the state-of-the-art model RITNet,respectively.展开更多
The Internet of Vehicles (IoV) has become an important direction in the field of intelligent transportation, in which vehicle positioning is a crucial part. SLAM (Simultaneous Localization and Mapping) technology play...The Internet of Vehicles (IoV) has become an important direction in the field of intelligent transportation, in which vehicle positioning is a crucial part. SLAM (Simultaneous Localization and Mapping) technology plays a crucial role in vehicle localization and navigation. Traditional Simultaneous Localization and Mapping (SLAM) systems are designed for use in static environments, and they can result in poor performance in terms of accuracy and robustness when used in dynamic environments where objects are in constant movement. To address this issue, a new real-time visual SLAM system called MG-SLAM has been developed. Based on ORB-SLAM2, MG-SLAM incorporates a dynamic target detection process that enables the detection of both known and unknown moving objects. In this process, a separate semantic segmentation thread is required to segment dynamic target instances, and the Mask R-CNN algorithm is applied on the Graphics Processing Unit (GPU) to accelerate segmentation. To reduce computational cost, only key frames are segmented to identify known dynamic objects. Additionally, a multi-view geometry method is adopted to detect unknown moving objects. The results demonstrate that MG-SLAM achieves higher precision, with an improvement from 0.2730 m to 0.0135 m in precision. Moreover, the processing time required by MG-SLAM is significantly reduced compared to other dynamic scene SLAM algorithms, which illustrates its efficacy in locating objects in dynamic scenes.展开更多
基金supported in part by the National Natural Science Foundation of China(No.31470714 and 61701105).
文摘Tree trunk instance segmentation is crucial for under-canopy unmanned aerial vehicles(UAVs)to autonomously extract standing tree stem attributes.Using cameras as sensors makes these UAVs compact and lightweight,facilitating safe and flexible navigation in dense forests.However,their limited onboard computational power makes real-time,image-based tree trunk segmentation challenging,emphasizing the urgent need for lightweight and efficient segmentation models.In this study,we present RT-Trunk,a model specifically designed for real-time tree trunk instance segmentation in complex forest environments.To ensure real-time performance,we selected SparseInst as the base framework.We incorporated ConvNeXt-T as the backbone to enhance feature extraction for tree trunks,thereby improving segmentation accuracy.We further integrate the lightweight convolutional block attention module(CBAM),enabling the model to focus on tree trunk features while suppressing irrelevant information,which leads to additional gains in segmentation accuracy.To enable RT-Trunk to operate effectively under diverse complex forest environments,we constructed a comprehensive dataset for training and testing by combining self-collected data with multiple public datasets covering different locations,seasons,weather conditions,tree species,and levels of forest clutter.Com-pared with the other tree trunk segmentation methods,the RT-Trunk method achieved an average precision of 91.4%and the fastest inference speed of 32.9 frames per second.Overall,the proposed RT-Trunk provides superior trunk segmentation performance that balances speed and accu-racy,making it a promising solution for supporting under-canopy UAVs in the autonomous extraction of standing tree stem attributes.The code for this work is available at https://github.com/NEFU CVRG/RT Trunk.
基金supported in part by the National Natural Science Foundation of China[Grant number 62471075]the Major Science and Technology Project Grant of the Chongqing Municipal Education Commission[Grant number KJZD-M202301901]Graduate Innovation Fund of Chongqing[gzlcx20253235].
文摘Semantic segmentation in street scenes is a crucial technology for autonomous driving to analyze the surrounding environment.In street scenes,issues such as high image resolution caused by a large viewpoints and differences in object scales lead to a decline in real-time performance and difficulties in multi-scale feature extraction.To address this,we propose a bilateral-branch real-time semantic segmentationmethod based on semantic information distillation(BSDNet)for street scene images.The BSDNet consists of a Feature Conversion Convolutional Block(FCB),a Semantic Information Distillation Module(SIDM),and a Deep Aggregation Atrous Convolution Pyramid Pooling(DASP).FCB reduces the semantic gap between the backbone and the semantic branch.SIDM extracts high-quality semantic information fromthe Transformer branch to reduce computational costs.DASP aggregates information lost in atrous convolutions,effectively capturingmulti-scale objects.Extensive experiments conducted on Cityscapes,CamVid,and ADE20K,achieving an accuracy of 81.7% Mean Intersection over Union(mIoU)at 70.6 Frames Per Second(FPS)on Cityscapes,demonstrate that our method achieves a better balance between accuracy and inference speed.
文摘Real-time semantic segmentation tasks place stringent demands on network inference speed,often requiring a reduction in network depth to decrease computational load.However,shallow networks tend to exhibit degradation in feature extraction completeness and inference accuracy.Therefore,balancing high performance with real-time requirements has become a critical issue in the study of real-time semantic segmentation.To address these challenges,this paper proposes a lightweight bilateral dual-residual network.By introducing a novel residual structure combined with feature extraction and fusion modules,the proposed network significantly enhances representational capacity while reducing computational costs.Specifically,an improved compound residual structure is designed to optimize the efficiency of information propagation and feature extraction.Furthermore,the proposed feature extraction and fusion module enables the network to better capture multi-scale information in images,improving the ability to detect both detailed and global semantic features.Experimental results on the publicly available Cityscapes dataset demonstrate that the proposed lightweight dual-branch network achieves outstanding performance while maintaining low computational complexity.In particular,the network achieved a mean Intersection over Union(mIoU)of 78.4%on the Cityscapes validation set,surpassing many existing semantic segmentation models.Additionally,in terms of inference speed,the network reached 74.5 frames per second when tested on an NVIDIA GeForce RTX 3090 GPU,significantly improving real-time performance.
基金provided by the Science Research Project of Hebei Education Department under grant No.BJK2024115.
文摘High-resolution remote sensing images(HRSIs)are now an essential data source for gathering surface information due to advancements in remote sensing data capture technologies.However,their significant scale changes and wealth of spatial details pose challenges for semantic segmentation.While convolutional neural networks(CNNs)excel at capturing local features,they are limited in modeling long-range dependencies.Conversely,transformers utilize multihead self-attention to integrate global context effectively,but this approach often incurs a high computational cost.This paper proposes a global-local multiscale context network(GLMCNet)to extract both global and local multiscale contextual information from HRSIs.A detail-enhanced filtering module(DEFM)is proposed at the end of the encoder to refine the encoder outputs further,thereby enhancing the key details extracted by the encoder and effectively suppressing redundant information.In addition,a global-local multiscale transformer block(GLMTB)is proposed in the decoding stage to enable the modeling of rich multiscale global and local information.We also design a stair fusion mechanism to transmit deep semantic information from deep to shallow layers progressively.Finally,we propose the semantic awareness enhancement module(SAEM),which further enhances the representation of multiscale semantic features through spatial attention and covariance channel attention.Extensive ablation analyses and comparative experiments were conducted to evaluate the performance of the proposed method.Specifically,our method achieved a mean Intersection over Union(mIoU)of 86.89%on the ISPRS Potsdam dataset and 84.34%on the ISPRS Vaihingen dataset,outperforming existing models such as ABCNet and BANet.
文摘Advanced traffic monitoring systems encounter substantial challenges in vehicle detection and classification due to the limitations of conventional methods,which often demand extensive computational resources and struggle with diverse data acquisition techniques.This research presents a novel approach for vehicle classification and recognition in aerial image sequences,integrating multiple advanced techniques to enhance detection accuracy.The proposed model begins with preprocessing using Multiscale Retinex(MSR)to enhance image quality,followed by Expectation-Maximization(EM)Segmentation for precise foreground object identification.Vehicle detection is performed using the state-of-the-art YOLOv10 framework,while feature extraction incorporates Maximally Stable Extremal Regions(MSER),Dense Scale-Invariant Feature Transform(Dense SIFT),and Zernike Moments Features to capture distinct object characteristics.Feature optimization is further refined through a Hybrid Swarm-based Optimization algorithm,ensuring optimal feature selection for improved classification performance.The final classification is conducted using a Vision Transformer,leveraging its robust learning capabilities for enhanced accuracy.Experimental evaluations on benchmark datasets,including UAVDT and the Unmanned Aerial Vehicle Intruder Dataset(UAVID),demonstrate the superiority of the proposed approach,achieving an accuracy of 94.40%on UAVDT and 93.57%on UAVID.The results highlight the efficacy of the model in significantly enhancing vehicle detection and classification in aerial imagery,outperforming existing methodologies and offering a statistically validated improvement for intelligent traffic monitoring systems compared to existing approaches.
文摘This systematic review aims to comprehensively examine and compare deep learning methods for brain tumor segmentation and classification using MRI and other imaging modalities,focusing on recent trends from 2022 to 2025.The primary objective is to evaluate methodological advancements,model performance,dataset usage,and existing challenges in developing clinically robust AI systems.We included peer-reviewed journal articles and highimpact conference papers published between 2022 and 2025,written in English,that proposed or evaluated deep learning methods for brain tumor segmentation and/or classification.Excluded were non-open-access publications,books,and non-English articles.A structured search was conducted across Scopus,Google Scholar,Wiley,and Taylor&Francis,with the last search performed in August 2025.Risk of bias was not formally quantified but considered during full-text screening based on dataset diversity,validation methods,and availability of performance metrics.We used narrative synthesis and tabular benchmarking to compare performance metrics(e.g.,accuracy,Dice score)across model types(CNN,Transformer,Hybrid),imaging modalities,and datasets.A total of 49 studies were included(43 journal articles and 6 conference papers).These studies spanned over 9 public datasets(e.g.,BraTS,Figshare,REMBRANDT,MOLAB)and utilized a range of imaging modalities,predominantly MRI.Hybrid models,especially ResViT and UNetFormer,consistently achieved high performance,with classification accuracy exceeding 98%and segmentation Dice scores above 0.90 across multiple studies.Transformers and hybrid architectures showed increasing adoption post2023.Many studies lacked external validation and were evaluated only on a few benchmark datasets,raising concerns about generalizability and dataset bias.Few studies addressed clinical interpretability or uncertainty quantification.Despite promising results,particularly for hybrid deep learning models,widespread clinical adoption remains limited due to lack of validation,interpretability concerns,and real-world deployment barriers.
基金supported by Institute of Information&Communications Technology Planning&Evaluation(IITP)under the Metaverse Support Program to Nurture the Best Talents(IITP-2024-RS-2023-00254529)grant funded by the Korea government(MSIT).
文摘Brain tumors require precise segmentation for diagnosis and treatment plans due to their complex morphology and heterogeneous characteristics.While MRI-based automatic brain tumor segmentation technology reduces the burden on medical staff and provides quantitative information,existing methodologies and recent models still struggle to accurately capture and classify the fine boundaries and diverse morphologies of tumors.In order to address these challenges and maximize the performance of brain tumor segmentation,this research introduces a novel SwinUNETR-based model by integrating a new decoder block,the Hierarchical Channel-wise Attention Decoder(HCAD),into a powerful SwinUNETR encoder.The HCAD decoder block utilizes hierarchical features and channelspecific attention mechanisms to further fuse information at different scales transmitted from the encoder and preserve spatial details throughout the reconstruction phase.Rigorous evaluations on the recent BraTS GLI datasets demonstrate that the proposed SwinHCAD model achieved superior and improved segmentation accuracy on both the Dice score and HD95 metrics across all tumor subregions(WT,TC,and ET)compared to baseline models.In particular,the rationale and contribution of the model design were clarified through ablation studies to verify the effectiveness of the proposed HCAD decoder block.The results of this study are expected to greatly contribute to enhancing the efficiency of clinical diagnosis and treatment planning by increasing the precision of automated brain tumor segmentation.
基金National Science and Technology Council,the Republic of China,under grants NSTC 113-2221-E-194-011-MY3 and Research Center on Artificial Intelligence and Sustainability,National Chung Cheng University under the research project grant titled“Generative Digital Twin System Design for Sustainable Smart City Development in Taiwan.
文摘Modern manufacturing processes have become more reliant on automation because of the accelerated transition from Industry 3.0 to Industry 4.0.Manual inspection of products on assembly lines remains inefficient,prone to errors and lacks consistency,emphasizing the need for a reliable and automated inspection system.Leveraging both object detection and image segmentation approaches,this research proposes a vision-based solution for the detection of various kinds of tools in the toolkit using deep learning(DL)models.Two Intel RealSense D455f depth cameras were arranged in a top down configuration to capture both RGB and depth images of the toolkits.After applying multiple constraints and enhancing them through preprocessing and augmentation,a dataset consisting of 3300 annotated RGB-D photos was generated.Several DL models were selected through a comprehensive assessment of mean Average Precision(mAP),precision-recall equilibrium,inference latency(target≥30 FPS),and computational burden,resulting in a preference for YOLO and Region-based Convolutional Neural Networks(R-CNN)variants over ViT-based models due to the latter’s increased latency and resource requirements.YOLOV5,YOLOV8,YOLOV11,Faster R-CNN,and Mask R-CNN were trained on the annotated dataset and evaluated using key performance metrics(Recall,Accuracy,F1-score,and Precision).YOLOV11 demonstrated balanced excellence with 93.0%precision,89.9%recall,and a 90.6%F1-score in object detection,as well as 96.9%precision,95.3%recall,and a 96.5%F1-score in instance segmentation with an average inference time of 25 ms per frame(≈40 FPS),demonstrating real-time performance.Leveraging these results,a YOLOV11-based windows application was successfully deployed in a real-time assembly line environment,where it accurately processed live video streams to detect and segment tools within toolkits,demonstrating its practical effectiveness in industrial automation.The application is capable of precisely measuring socket dimensions by utilising edge detection techniques on YOLOv11 segmentation masks,in addition to detection and segmentation.This makes it possible to do specification-level quality control right on the assembly line,which improves the ability to examine things in real time.The implementation is a big step forward for intelligent manufacturing in the Industry 4.0 paradigm.It provides a scalable,efficient,and accurate way to do automated inspection and dimensional verification activities.
基金supported in part by the National Key RD Program of China (2021YFF0602104-2,2020YFB1804604)in part by the 2020 Industrial Internet Innovation and Development Project from Ministry of Industry and Information Technology of Chinain part by the Fundamental Research Fund for the Central Universities (30918012204,30920041112).
文摘The application of unmanned driving in the Internet of Things is one of the concrete manifestations of the application of artificial intelligence technology.Image semantic segmentation can help the unmanned driving system by achieving road accessibility analysis.Semantic segmentation is also a challenging technology for image understanding and scene parsing.We focused on the challenging task of real-time semantic segmentation in this paper.In this paper,we proposed a novel fast architecture for real-time semantic segmentation named DuFNet.Starting from the existing work of Bilateral Segmentation Network(BiSeNet),DuFNet proposes a novel Semantic Information Flow(SIF)structure for context information and a novel Fringe Information Flow(FIF)structure for spatial information.We also proposed two kinds of SIF with cascaded and paralleled structures,respectively.The SIF encodes the input stage by stage in the ResNet18 backbone and provides context information for the feature fusionmodule.Features from previous stages usually contain rich low-level details but high-level semantics for later stages.Themultiple convolutions embed in Parallel SIF aggregate the corresponding features among different stages and generate a powerful global context representation with less computational cost.The FIF consists of a pooling layer and an upsampling operator followed by projection convolution layer.The concise component provides more spatial details for the network.Compared with BiSeNet,our work achieved faster speed and comparable performance with 72.34%mIoU accuracy and 78 FPS on Cityscapes Dataset based on the ResNet18 backbone.
基金supported by the Natural Science Foundation of Guizhou Province(Grant Number:20161054)Joint Natural Science Foundation of Guizhou Province(Grant Number:LH20177226)+1 种基金2017 Special Project of New Academic Talent Training and Innovation Exploration of Guizhou University(Grant Number:20175788)The National Natural Science Foundation of China under Grant No.12205062.
文摘Autonomous driving technology has made a lot of outstanding achievements with deep learning,and the vehicle detection and classification algorithm has become one of the critical technologies of autonomous driving systems.The vehicle instance segmentation can perform instance-level semantic parsing of vehicle information,which is more accurate and reliable than object detection.However,the existing instance segmentation algorithms still have the problems of poor mask prediction accuracy and low detection speed.Therefore,this paper proposes an advanced real-time instance segmentation model named FIR-YOLACT,which fuses the ICIoU(Improved Complete Intersection over Union)and Res2Net for the YOLACT algorithm.Specifically,the ICIoU function can effectively solve the degradation problem of the original CIoU loss function,and improve the training convergence speed and detection accuracy.The Res2Net module fused with the ECA(Efficient Channel Attention)Net is added to the model’s backbone network,which improves the multi-scale detection capability and mask prediction accuracy.Furthermore,the Cluster NMS(Non-Maximum Suppression)algorithm is introduced in the model’s bounding box regression to enhance the performance of detecting similarly occluded objects.The experimental results demonstrate the superiority of FIR-YOLACT to the based methods and the effectiveness of all components.The processing speed reaches 28 FPS,which meets the demands of real-time vehicle instance segmentation.
基金funded by Anhui Provincial Natural Science Foundation(No.2208085ME128)the Anhui University-Level Special Project of Anhui University of Science and Technology(No.XCZX2021-01)+1 种基金the Research and the Development Fund of the Institute of Environmental Friendly Materials and Occupational Health,Anhui University of Science and Technology(No.ALW2022YF06)Anhui Province New Era Education Quality Project(Graduate Education)(No.2022xscx073).
文摘The real-time detection and instance segmentation of strawberries constitute fundamental components in the development of strawberry harvesting robots.Real-time identification of strawberries in an unstructured envi-ronment is a challenging task.Current instance segmentation algorithms for strawberries suffer from issues such as poor real-time performance and low accuracy.To this end,the present study proposes an Efficient YOLACT(E-YOLACT)algorithm for strawberry detection and segmentation based on the YOLACT framework.The key enhancements of the E-YOLACT encompass the development of a lightweight attention mechanism,pyramid squeeze shuffle attention(PSSA),for efficient feature extraction.Additionally,an attention-guided context-feature pyramid network(AC-FPN)is employed instead of FPN to optimize the architecture’s performance.Furthermore,a feature-enhanced model(FEM)is introduced to enhance the prediction head’s capabilities,while efficient fast non-maximum suppression(EF-NMS)is devised to improve non-maximum suppression.The experimental results demonstrate that the E-YOLACT achieves a Box-mAP and Mask-mAP of 77.9 and 76.6,respectively,on the custom dataset.Moreover,it exhibits an impressive category accuracy of 93.5%.Notably,the E-YOLACT also demonstrates a remarkable real-time detection capability with a speed of 34.8 FPS.The method proposed in this article presents an efficient approach for the vision system of a strawberry-picking robot.
基金co-supported by the National Natural Science Foundation of China(Nos.U1833117 and 61806015)the National Key Research and Development Program of China(No.2017YFB0503402)。
文摘The aircraft system has recently gained its reputation as a reliable and efficient tool for sensing and parsing aerial scenes.However,accurate and fast semantic segmentation of highresolution aerial images for remote sensing applications is still facing three challenges:the requirements for limited processing resources and low-latency operations based on aerial platforms,the balance between high accuracy and real-time efficiency for model performance,and the confusing objects with large intra-class variations and small inter-class differences in high-resolution aerial images.To address these issues,a lightweight and dual-path deep convolutional architecture,namely Aerial Bilateral Segmentation Network(Aerial-Bi Se Net),is proposed to perform realtime segmentation on high-resolution aerial images with favorable accuracy.Specifically,inspired by the receptive field concept in human visual systems,Receptive Field Module(RFM)is proposed to encode rich multi-scale contextual information.Based on channel attention mechanism,two novel modules,called Feature Attention Module(FAM)and Channel Attention based Feature Fusion Module(CAFFM)respectively,are proposed to refine and combine features effectively to boost the model performance.Aerial-Bi Se Net is evaluated on the Potsdam and Vaihingen datasets,where leading performance is reported compared with other state-of-the-art models,in terms of both accuracy and efficiency.
文摘As the field of autonomous driving evolves, real-time semantic segmentation has become a crucial part of computer vision tasks. However, most existing methods use lightweight convolution to reduce the computational effort, resulting in lower accuracy. To address this problem, we construct TBANet, a network with an encoder-decoder structure for efficient feature extraction. In the encoder part, the TBA module is designed to extract details and the ETBA module is used to learn semantic representations in a high-dimensional space. In the decoder part, we design a combination of multiple upsampling methods to aggregate features with less computational overhead. We validate the efficiency of TBANet on the Cityscapes dataset. It achieves 75.1% mean Intersection over Union(mIoU) with only 2.07 million parameters and can reach 90.3 Frames Per Second(FPS).
基金supported by the National Natural Science Foundation of China(61233010 61305106)+2 种基金the Shanghai Natural Science Foundation(17ZR1409700 18ZR1415300)the basic research project of Shanghai Municipal Science and Technology Commission(16JC1400900)
文摘This paper concerns the problem of object segmentation in real-time for picking system. A region proposal method inspired by human glance based on the convolutional neural network is proposed to select promising regions, allowing more processing is reserved only for these regions. The speed of object segmentation is significantly improved by the region proposal method.By the combination of the region proposal method based on the convolutional neural network and superpixel method, the category and location information can be used to segment objects and image redundancy is significantly reduced. The processing time is reduced considerably by this to achieve the real time. Experiments show that the proposed method can segment the interested target object in real time on an ordinary laptop.
基金The National Natural Science Foundation of China (62176048)provided funding for this research.
文摘The interpretability of deep learning models has emerged as a compelling area in artificial intelligence research.The safety criteria for medical imaging are highly stringent,and models are required for an explanation.However,existing convolutional neural network solutions for left ventricular segmentation are viewed in terms of inputs and outputs.Thus,the interpretability of CNNs has come into the spotlight.Since medical imaging data are limited,many methods to fine-tune medical imaging models that are popular in transfer models have been built using massive public Image Net datasets by the transfer learning method.Unfortunately,this generates many unreliable parameters and makes it difficult to generate plausible explanations from these models.In this study,we trained from scratch rather than relying on transfer learning,creating a novel interpretable approach for autonomously segmenting the left ventricle with a cardiac MRI.Our enhanced GPU training system implemented interpretable global average pooling for graphics using deep learning.The deep learning tasks were simplified.Simplification included data management,neural network architecture,and training.Our system monitored and analyzed the gradient changes of different layers with dynamic visualizations in real-time and selected the optimal deployment model.Our results demonstrated that the proposed method was feasible and efficient:the Dice coefficient reached 94.48%,and the accuracy reached 99.7%.It was found that no current transfer learning models could perform comparably to the ImageNet transfer learning architectures.This model is lightweight and more convenient to deploy on mobile devices than transfer learning models.
文摘While traditional Convolutional Neural Network(CNN)-based semantic segmentation methods have proven effective,they often encounter significant computational challenges due to the requirement for dense pixel-level predictions,which complicates real-time implementation.To address this,we introduce an advanced real-time semantic segmentation strategy specifically designed for autonomous driving,utilizing the capabilities of Visual Transformers.By leveraging the self-attention mechanism inherent in Visual Transformers,our method enhances global contextual awareness,refining the representation of each pixel in relation to the overall scene.This enhancement is critical for quickly and accurately interpreting the complex elements within driving sce-narios—a fundamental need for autonomous vehicles.Our experiments conducted on the DriveSeg autonomous driving dataset indicate that our model surpasses traditional segmentation methods,achieving a significant 4.5%improvement in Mean Intersection over Union(mIoU)while maintaining real-time responsiveness.This paper not only underscores the potential for optimized semantic segmentation but also establishes a promising direction for real-time processing in autonomous navigation systems.Future work will focus on integrating this technique with other perception modules in autonomous driving to further improve the robustness and efficiency of self-driving perception frameworks,thereby opening new pathways for research and practical applications in scenarios requiring rapid and precise decision-making capabilities.Further experimentation and adaptation of this model could lead to broader implications for the fields of machine learning and computer vision,particularly in enhancing the interaction between automated systems and their dynamic environments.
文摘Background Generally, it is difficult to obtain accurate pose and depth for a non-rigid moving object from a single RGB camera to create augmented reality (AR). In this study, we build an augmented reality system from a single RGB camera for a non-rigid moving human by accurately computing pose and depth, for which two key tasks are segmentation and monocular Simultaneous Localization and Mapping (SLAM). Most existing monocular SLAM systems are designed for static scenes, while in this AR system, the human body is always moving and non-rigid. Methods In order to make the SLAM system suitable for a moving human, we first segment the rigid part of the human in each frame. A segmented moving body part can be regarded as a static object, and the relative motions between each moving body part and the camera can be considered the motion of the camera. Typical SLAM systems designed for static scenes can then be applied. In the segmentation step of this AR system, we first employ the proposed BowtieNet, which adds the atrous spatial pyramid pooling (ASPP) of DeepLab between the encoder and decoder of SegNet to segment the human in the original frame, and then we use color information to extract the face from the segmented human area. Results Based on the human segmentation results and a monocular SLAM, this system can change the video background and add a virtual object to humans. Conclusions The experiments on the human image segmentation datasets show that BowtieNet obtains state-of-the-art human image segmentation performance and enough speed for real-time segmentation. The experiments on videos show that the proposed AR system can robustly add a virtual object to humans and can accurately change the video background.
基金supported by National Key Research and Development Program(No.2022YFE0112400)National Natural Science Foundation of China(No.21706096)Natural Science Foundation of Jiangsu Province(No.BK20160162).
文摘Instance segmentation plays an important role in image processing.The Deep Snake algorithm based on contour iteration deforms an initial bounding box to an instance contour end-to-end,which can improve the performance of instance segmentation,but has defects such as slow segmentation speed and sub-optimal initial contour.To solve these problems,a real-time instance segmentation algorithm based on contour learning was proposed.Firstly,ShuffleNet V2 was used as backbone network,and the receptive field of the model was expanded by using a 5×5 convolution kernel.Secondly,a lightweight up-sampling module,multi-stage aggregation(MSA),performs residual fusion of multi-layer features,which not only improves segmentation speed,but also extracts effective features more comprehensively.Thirdly,a contour initialization method for network learning was designed,and a global contour feature aggregation mechanism was used to return a coarse contour,which solves the problem of excessive error between manually initialized contour and real contour.Finally,the Snake deformation module was used to iteratively optimize the coarse contour to obtain the final instance contour.The experimental results showed that the proposed method improved the instance segmentation accuracy on semantic boundaries dataset(SBD),Cityscapes and Kins datasets,and the average precision reached 55.8 on the SBD;Compared with Deep Snake,the model parameters were reduced by 87.2%,calculation amount was reduced by 78.3%,and segmentation speed reached 39.8 frame·s^(−1) when instance segmentation was performed on an image with a size of 512×512 pixels on a 2080Ti GPU.The proposed method can reduce resource consumption,realize instance segmentation tasks quickly and accurately,and therefore is more suitable for embedded platforms with limited resources.
基金supported by the HFIPS Director’s Foundation(YZJJ202207-TS),the National Natural Science Foundation of China(82371931)the Natural Science Foundation of Anhui Province(2008085MC69)+3 种基金the Natural Science Foundation of Hefei City(2021033)the General Scientific Research Project of Anhui Provincial Health Commission(AHWJ2021b150)the Collaborative Innovation Program of Hefei Science Center,CAS(2021HSC-CIP013)the Anhui Province Key Research and Development Project(202204295107020004).
文摘Semantic segmentation of eye images is a complex task with important applications in human–computer interaction,cognitive science,and neuroscience.Achieving real-time,accurate,and robust segmentation algorithms is crucial for computationally limited portable devices such as augmented reality and virtual reality.With the rapid advancements in deep learning,many network models have been developed specifically for eye image segmentation.Some methods divide the segmentation process into multiple stages to achieve model parameter miniaturization while enhancing output through post processing techniques to improve segmentation accuracy.These approaches significantly increase the inference time.Other networks adopt more complex encoding and decoding modules to achieve end-to-end output,which requires substantial computation.Therefore,balancing the model’s size,accuracy,and computational complexity is essential.To address these challenges,we propose a lightweight asymmetric UNet architecture and a projection loss function.We utilize ResNet-3 layer blocks to enhance feature extraction efficiency in the encoding stage.In the decoding stage,we employ regular convolutions and skip connections to upscale the feature maps from the latent space to the original image size,balancing the model size and segmentation accuracy.In addition,we leverage the geometric features of the eye region and design a projection loss function to further improve the segmentation accuracy without adding any additional inference computational cost.We validate our approach on the OpenEDS2019 dataset for virtual reality and achieve state-of-the-art performance with 95.33%mean intersection over union(mIoU).Our model has only 0.63M parameters and 350 FPS,which are 68%and 200%of the state-of-the-art model RITNet,respectively.
基金funded by the Natural Science Foundation of the Jiangsu Higher Education Institutions of China(grant number 22KJD440001)Changzhou Science&Technology Program(grant number CJ20220232).
文摘The Internet of Vehicles (IoV) has become an important direction in the field of intelligent transportation, in which vehicle positioning is a crucial part. SLAM (Simultaneous Localization and Mapping) technology plays a crucial role in vehicle localization and navigation. Traditional Simultaneous Localization and Mapping (SLAM) systems are designed for use in static environments, and they can result in poor performance in terms of accuracy and robustness when used in dynamic environments where objects are in constant movement. To address this issue, a new real-time visual SLAM system called MG-SLAM has been developed. Based on ORB-SLAM2, MG-SLAM incorporates a dynamic target detection process that enables the detection of both known and unknown moving objects. In this process, a separate semantic segmentation thread is required to segment dynamic target instances, and the Mask R-CNN algorithm is applied on the Graphics Processing Unit (GPU) to accelerate segmentation. To reduce computational cost, only key frames are segmented to identify known dynamic objects. Additionally, a multi-view geometry method is adopted to detect unknown moving objects. The results demonstrate that MG-SLAM achieves higher precision, with an improvement from 0.2730 m to 0.0135 m in precision. Moreover, the processing time required by MG-SLAM is significantly reduced compared to other dynamic scene SLAM algorithms, which illustrates its efficacy in locating objects in dynamic scenes.
