Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware reso...Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.展开更多
Human pose estimation aims to localize the body joints from image or video data.With the development of deeplearning,pose estimation has become a hot research topic in the field of computer vision.In recent years,huma...Human pose estimation aims to localize the body joints from image or video data.With the development of deeplearning,pose estimation has become a hot research topic in the field of computer vision.In recent years,humanpose estimation has achieved great success in multiple fields such as animation and sports.However,to obtainaccurate positioning results,existing methods may suffer from large model sizes,a high number of parameters,and increased complexity,leading to high computing costs.In this paper,we propose a new lightweight featureencoder to construct a high-resolution network that reduces the number of parameters and lowers the computingcost.We also introduced a semantic enhancement module that improves global feature extraction and networkperformance by combining channel and spatial dimensions.Furthermore,we propose a dense connected spatialpyramid pooling module to compensate for the decrease in image resolution and information loss in the network.Finally,ourmethod effectively reduces the number of parameters and complexitywhile ensuring high performance.Extensive experiments show that our method achieves a competitive performance while dramatically reducing thenumber of parameters,and operational complexity.Specifically,our method can obtain 89.9%AP score on MPIIVAL,while the number of parameters and the complexity of operations were reduced by 41%and 36%,respectively.展开更多
The employment of deep convolutional neural networks has recently contributed to significant progress in single image super-resolution(SISR)research.However,the high computational demands of most SR techniques hinder ...The employment of deep convolutional neural networks has recently contributed to significant progress in single image super-resolution(SISR)research.However,the high computational demands of most SR techniques hinder their applicability to edge devices,despite their satisfactory reconstruction performance.These methods commonly use standard convolutions,which increase the convolutional operation cost of the model.In this paper,a lightweight Partial Separation and Multiscale Fusion Network(PSMFNet)is proposed to alleviate this problem.Specifically,this paper introduces partial convolution(PConv),which reduces the redundant convolution operations throughout the model by separating some of the features of an image while retaining features useful for image reconstruction.Additionally,it is worth noting that the existing methods have not fully utilized the rich feature information,leading to information loss,which reduces the ability to learn feature representations.Inspired by self-attention,this paper develops a multiscale feature fusion block(MFFB),which can better utilize the non-local features of an image.MFFB can learn long-range dependencies from the spatial dimension and extract features from the channel dimension,thereby obtaining more comprehensive and rich feature information.As the role of the MFFB is to capture rich global features,this paper further introduces an efficient inverted residual block(EIRB)to supplement the local feature extraction ability of PSMFNet.A comprehensive analysis of the experimental results shows that PSMFNet maintains a better performance with fewer parameters than the state-of-the-art models.展开更多
In minimally invasive surgery,endoscopes or laparoscopes equipped with miniature cameras and tools are used to enter the human body for therapeutic purposes through small incisions or natural cavities.However,in clini...In minimally invasive surgery,endoscopes or laparoscopes equipped with miniature cameras and tools are used to enter the human body for therapeutic purposes through small incisions or natural cavities.However,in clinical operating environments,endoscopic images often suffer from challenges such as low texture,uneven illumination,and non-rigid structures,which affect feature observation and extraction.This can severely impact surgical navigation or clinical diagnosis due to missing feature points in endoscopic images,leading to treatment and postoperative recovery issues for patients.To address these challenges,this paper introduces,for the first time,a Cross-Channel Multi-Modal Adaptive Spatial Feature Fusion(ASFF)module based on the lightweight architecture of EfficientViT.Additionally,a novel lightweight feature extraction and matching network based on attention mechanism is proposed.This network dynamically adjusts attention weights for cross-modal information from grayscale images and optical flow images through a dual-branch Siamese network.It extracts static and dynamic information features ranging from low-level to high-level,and from local to global,ensuring robust feature extraction across different widths,noise levels,and blur scenarios.Global and local matching are performed through a multi-level cascaded attention mechanism,with cross-channel attention introduced to simultaneously extract low-level and high-level features.Extensive ablation experiments and comparative studies are conducted on the HyperKvasir,EAD,M2caiSeg,CVC-ClinicDB,and UCL synthetic datasets.Experimental results demonstrate that the proposed network improves upon the baseline EfficientViT-B3 model by 75.4%in accuracy(Acc),while also enhancing runtime performance and storage efficiency.When compared with the complex DenseDescriptor feature extraction network,the difference in Acc is less than 7.22%,and IoU calculation results on specific datasets outperform complex dense models.Furthermore,this method increases the F1 score by 33.2%and accelerates runtime by 70.2%.It is noteworthy that the speed of CMMCAN surpasses that of comparative lightweight models,with feature extraction and matching performance comparable to existing complex models but with faster speed and higher cost-effectiveness.展开更多
In pursuit of cost-effective manufacturing,enterprises are increasingly adopting the practice of utilizing recycled semiconductor chips.To ensure consistent chip orientation during packaging,a circular marker on the f...In pursuit of cost-effective manufacturing,enterprises are increasingly adopting the practice of utilizing recycled semiconductor chips.To ensure consistent chip orientation during packaging,a circular marker on the front side is employed for pin alignment following successful functional testing.However,recycled chips often exhibit substantial surface wear,and the identification of the relatively small marker proves challenging.Moreover,the complexity of generic target detection algorithms hampers seamless deployment.Addressing these issues,this paper introduces a lightweight YOLOv8s-based network tailored for detecting markings on recycled chips,termed Van-YOLOv8.Initially,to alleviate the influence of diminutive,low-resolution markings on the precision of deep learning models,we utilize an upscaling approach for enhanced resolution.This technique relies on the Super-Resolution Generative Adversarial Network with Extended Training(SRGANext)network,facilitating the reconstruction of high-fidelity images that align with input specifications.Subsequently,we replace the original YOLOv8smodel’s backbone feature extraction network with the lightweight VanillaNetwork(VanillaNet),simplifying the branch structure to reduce network parameters.Finally,a Hybrid Attention Mechanism(HAM)is implemented to capture essential details from input images,improving feature representation while concurrently expediting model inference speed.Experimental results demonstrate that the Van-YOLOv8 network outperforms the original YOLOv8s on a recycled chip dataset in various aspects.Significantly,it demonstrates superiority in parameter count,computational intricacy,precision in identifying targets,and speed when compared to certain prevalent algorithms in the current landscape.The proposed approach proves promising for real-time detection of recycled chips in practical factory settings.展开更多
Automatic crack detection of cement pavement chiefly benefits from the rapid development of deep learning,with convolutional neural networks(CNN)playing an important role in this field.However,as the performance of cr...Automatic crack detection of cement pavement chiefly benefits from the rapid development of deep learning,with convolutional neural networks(CNN)playing an important role in this field.However,as the performance of crack detection in cement pavement improves,the depth and width of the network structure are significantly increased,which necessitates more computing power and storage space.This limitation hampers the practical implementation of crack detection models on various platforms,particularly portable devices like small mobile devices.To solve these problems,we propose a dual-encoder-based network architecture that focuses on extracting more comprehensive fracture feature information and combines cross-fusion modules and coordinated attention mechanisms formore efficient feature fusion.Firstly,we use small channel convolution to construct shallow feature extractionmodule(SFEM)to extract low-level feature information of cracks in cement pavement images,in order to obtainmore information about cracks in the shallowfeatures of images.In addition,we construct large kernel atrous convolution(LKAC)to enhance crack information,which incorporates coordination attention mechanism for non-crack information filtering,and large kernel atrous convolution with different cores,using different receptive fields to extract more detailed edge and context information.Finally,the three-stage feature map outputs from the shallow feature extraction module is cross-fused with the two-stage feature map outputs from the large kernel atrous convolution module,and the shallow feature and detailed edge feature are fully fused to obtain the final crack prediction map.We evaluate our method on three public crack datasets:DeepCrack,CFD,and Crack500.Experimental results on theDeepCrack dataset demonstrate the effectiveness of our proposed method compared to state-of-the-art crack detection methods,which achieves Precision(P)87.2%,Recall(R)87.7%,and F-score(F1)87.4%.Thanks to our lightweight crack detectionmodel,the parameter count of the model in real-world detection scenarios has been significantly reduced to less than 2M.This advancement also facilitates technical support for portable scene detection.展开更多
There is a problem of real-time detection difficulty in road surface damage detection. This paper proposes an improved lightweight model based on you only look once version 5(YOLOv5). Firstly, this paper fully utilize...There is a problem of real-time detection difficulty in road surface damage detection. This paper proposes an improved lightweight model based on you only look once version 5(YOLOv5). Firstly, this paper fully utilized the convolutional neural network(CNN) + ghosting bottleneck(G_bneck) architecture to reduce redundant feature maps. Afterwards, we upgraded the original upsampling algorithm to content-aware reassembly of features(CARAFE) and increased the receptive field. Finally, we replaced the spatial pyramid pooling fast(SPPF) module with the basic receptive field block(Basic RFB) pooling module and added dilated convolution. After comparative experiments, we can see that the number of parameters and model size of the improved algorithm in this paper have been reduced by nearly half compared to the YOLOv5s. The frame rate per second(FPS) has been increased by 3.25 times. The mean average precision(m AP@0.5: 0.95) has increased by 8%—17% compared to other lightweight algorithms.展开更多
Current you only look once(YOLO)-based algorithm model is facing the challenge of overwhelming parameters and calculation complexity under the printed circuit board(PCB)defect detection application scenario.In order t...Current you only look once(YOLO)-based algorithm model is facing the challenge of overwhelming parameters and calculation complexity under the printed circuit board(PCB)defect detection application scenario.In order to solve this problem,we propose a new method,which combined the lightweight network mobile vision transformer(Mobile Vi T)with the convolutional block attention module(CBAM)mechanism and the new regression loss function.This method needed less computation resources,making it more suitable for embedded edge detection devices.Meanwhile,the new loss function improved the positioning accuracy of the bounding box and enhanced the robustness of the model.In addition,experiments on public datasets demonstrate that the improved model achieves an average accuracy of 87.9%across six typical defect detection tasks,while reducing computational costs by nearly 90%.It significantly reduces the model's computational requirements while maintaining accuracy,ensuring reliable performance for edge deployment.展开更多
Mineral identification is foundational to geological survey research,mineral resource exploration,and mining engineering.Considering the diversity of mineral types and the challenge of achieving high recognition accur...Mineral identification is foundational to geological survey research,mineral resource exploration,and mining engineering.Considering the diversity of mineral types and the challenge of achieving high recognition accuracy for similar features,this study introduces a mineral detection method based on YOLOv8-SBI.This work enhances feature extraction by integrating spatial pyramid pooling-fast(SPPF)with the simplified self-attention module(SimAM),significantly improving the precision of mineral feature detection.In the feature fusion network,a weighted bidirectional feature pyramid network is employed for advanced cross-channel feature integration,effectively reducing feature redundancy.Additionally,Inner-Intersection Over Union(InnerIOU)is used as the loss function to improve the average quality localization performance of anchor boxes.Experimental results show that the YOLOv8-SBI model achieves an accuracy of 67.9%,a recall of 74.3%,a mAP@0.5 of 75.8%,and a mAP@0.5:0.95 of 56.7%,with a real-time detection speed of 244.2 frames per second.Compared to YOLOv8,YOLOv8-SBI demonstrates a significant improvement with 15.4%increase in accuracy,28.5%increase in recall,and increases of 28.1%and 20.9%in mAP@0.5 and mAP@0.5:0.95,respectively.Furthermore,relative to other models,such as YOLOv3,YOLOv5,YOLOv6,YOLOv8,YOLOv9,and YOLOv10,YOLOv8-SBI has a smaller parameter size of only 3.01×10^(6).This highlights the optimal balance between detection accuracy and speed,thereby offering robust technical support for intelligent mineral classification.展开更多
Accurately identifying crop pests and diseases ensures agricultural productivity and safety.Although current YOLO-based detection models offer real-time capabilities,their conventional convolutional layers involve hig...Accurately identifying crop pests and diseases ensures agricultural productivity and safety.Although current YOLO-based detection models offer real-time capabilities,their conventional convolutional layers involve high computational redundancy and a fixed receptive field,making it challenging to capture local details and global semantics in complex scenarios simultaneously.This leads to significant issues like missed detections of small targets and heightened sensitivity to background interference.To address these challenges,this paper proposes a lightweight adaptive detection network—StarSpark-AdaptiveNet(SSANet),which optimizes features through a dual-module collaborative mechanism.Specifically,the StarNet module utilizes Depthwise separable convolutions(DW-Conv)and dynamic star operations to establish multi-stage feature extraction pathways,enhancing local detail perception within a lightweight framework.Moreover,the Multi-scale Adaptive Spatial Attention Gate(MASAG)module integrates cross-layer feature fusion and dynamic weight allocation to capture multi-scale global contextual information,effectively suppressing background noise.These modules jointly form a“local enhancement-global calibration”bidirectional optimization mechanism,significantly improving the model’s adaptability to complex disease patterns.Furthermore,the proposed Scale-based Dynamic Loss(SD Loss)dynamically adjusts the weight of scale and localization losses,improving regression stability and localization accuracy,especially for small targets.Experiments on the eggplant fruit disease dataset demonstrate that SSANet achieves an mAP50 of 83.9%and a detection speed of 273.5 FPS with only 2.11 M parameters and 5.1 GFLOPs computational cost,outperforming the baseline YOLO11 model by reducing parameters by 18.1%,increasing mAP50 by 1.3%,and improving inference speed by 9.1%.Ablation studies further confirm the effectiveness and complementarity of the modules.SSANet offers a high-accuracy,low-cost solution suitable for real-time pest and disease detection in crops,facilitating edge device deployment and promoting precision agriculture.展开更多
This paper introduces a lightweight remote sensing image dehazing network called multidimensional weight regulation network(MDWR-Net), which addresses the high computational cost of existing methods. Previous works, o...This paper introduces a lightweight remote sensing image dehazing network called multidimensional weight regulation network(MDWR-Net), which addresses the high computational cost of existing methods. Previous works, often based on the encoder-decoder structure and utilizing multiple upsampling and downsampling layers, are computationally expensive. To improve efficiency, the paper proposes two modules: the efficient spatial resolution recovery module(ESRR) for upsampling and the efficient depth information augmentation module(EDIA) for downsampling.These modules not only reduce model complexity but also enhance performance. Additionally, the partial feature weight learning module(PFWL) is introduced to reduce the computational burden by applying weight learning across partial dimensions, rather than using full-channel convolution.To overcome the limitations of convolutional neural networks(CNN)-based networks, the haze distribution index transformer(HDIT) is integrated into the decoder. We also propose the physicalbased non-adjacent feature fusion module(PNFF), which leverages the atmospheric scattering model to improve generalization of our MDWR-Net. The MDWR-Net achieves superior dehazing performance with a computational cost of just 2.98×10^(9) multiply-accumulate operations(MACs),which is less than one-tenth of previous methods. Experimental results validate its effectiveness in balancing performance and computational efficiency.展开更多
In recent years,the country has spent significant workforce and material resources to prevent traffic accidents,particularly those caused by fatigued driving.The current studies mainly concentrate on driver physiologi...In recent years,the country has spent significant workforce and material resources to prevent traffic accidents,particularly those caused by fatigued driving.The current studies mainly concentrate on driver physiological signals,driving behavior,and vehicle information.However,most of the approaches are computationally intensive and inconvenient for real-time detection.Therefore,this paper designs a network that combines precision,speed and lightweight and proposes an algorithm for facial fatigue detection based on multi-feature fusion.Specifically,the face detection model takes YOLOv8(You Only Look Once version 8)as the basic framework,and replaces its backbone network with MobileNetv3.To focus on the significant regions in the image,CPCA(Channel Prior Convolution Attention)is adopted to enhance the network’s capacity for feature extraction.Meanwhile,the network training phase employs the Focal-EIOU(Focal and Efficient Intersection Over Union)loss function,which makes the network lightweight and increases the accuracy of target detection.Ultimately,the Dlib toolkit was employed to annotate 68 facial feature points.This study established an evaluation metric for facial fatigue and developed a novel fatigue detection algorithm to assess the driver’s condition.A series of comparative experiments were carried out on the self-built dataset.The suggested method’s mAP(mean Average Precision)values for object detection and fatigue detection are 96.71%and 95.75%,respectively,as well as the detection speed is 47 FPS(Frames Per Second).This method can balance the contradiction between computational complexity and model accuracy.Furthermore,it can be transplanted to NVIDIA Jetson Orin NX and quickly detect the driver’s state while maintaining a high degree of accuracy.It contributes to the development of automobile safety systems and reduces the occurrence of traffic accidents.展开更多
Accurate cloud classification plays a crucial role in aviation safety,climate monitoring,and localized weather forecasting.Current research has been focusing on machine learning techniques,particularly deep learning b...Accurate cloud classification plays a crucial role in aviation safety,climate monitoring,and localized weather forecasting.Current research has been focusing on machine learning techniques,particularly deep learning based model,for the types identification.However,traditional approaches such as convolutional neural networks(CNNs)encounter difficulties in capturing global contextual information.In addition,they are computationally expensive,which restricts their usability in resource-limited environments.To tackle these issues,we present the Cloud Vision Transformer(CloudViT),a lightweight model that integrates CNNs with Transformers.The integration enables an effective balance between local and global feature extraction.To be specific,CloudViT comprises two innovative modules:Feature Extraction(E_Module)and Downsampling(D_Module).These modules are able to significantly reduce the number of model parameters and computational complexity while maintaining translation invariance and enhancing contextual comprehension.Overall,the CloudViT includes 0.93×10^(6)parameters,which decreases more than ten times compared to the SOTA(State-of-the-Art)model CloudNet.Comprehensive evaluations conducted on the HBMCD and SWIMCAT datasets showcase the outstanding performance of CloudViT.It achieves classification accuracies of 98.45%and 100%,respectively.Moreover,the efficiency and scalability of CloudViT make it an ideal candidate for deployment inmobile cloud observation systems,enabling real-time cloud image classification.The proposed hybrid architecture of CloudViT offers a promising approach for advancing ground-based cloud image classification.It holds significant potential for both optimizing performance and facilitating practical deployment scenarios.展开更多
Vehicle re-identification involves matching images of vehicles across varying camera views.The diversity of camera locations along different roadways leads to significant intra-class variation and only minimal inter-c...Vehicle re-identification involves matching images of vehicles across varying camera views.The diversity of camera locations along different roadways leads to significant intra-class variation and only minimal inter-class similarity in the collected vehicle images,which increases the complexity of re-identification tasks.To tackle these challenges,this study proposes AG-GCN(Attention-Guided Graph Convolutional Network),a novel framework integrating several pivotal components.Initially,AG-GCN embeds a lightweight attention module within the ResNet-50 structure to learn feature weights automatically,thereby improving the representation of vehicle features globally by highlighting salient features and suppressing extraneous ones.Moreover,AG-GCN adopts a graph-based structure to encapsulate deep local features.A graph convolutional network then amalgamates these features to understand the relationships among vehicle-related characteristics.Subsequently,we amalgamate feature maps from both the attention and graph-based branches for a more comprehensive representation of vehicle features.The framework then gauges feature similarities and ranks them,thus enhancing the accuracy of vehicle re-identification.Comprehensive qualitative and quantitative analyses on two publicly available datasets verify the efficacy of AG-GCN in addressing intra-class and inter-class variability issues.展开更多
The tradeoff between efficiency and model size of the convolutional neural network(CNN)is an essential issue for applications of CNN-based algorithms to diverse real-world tasks.Although deep learning-based methods ha...The tradeoff between efficiency and model size of the convolutional neural network(CNN)is an essential issue for applications of CNN-based algorithms to diverse real-world tasks.Although deep learning-based methods have achieved significant improvements in image super-resolution(SR),current CNNbased techniques mainly contain massive parameters and a high computational complexity,limiting their practical applications.In this paper,we present a fast and lightweight framework,named weighted multi-scale residual network(WMRN),for a better tradeoff between SR performance and computational efficiency.With the modified residual structure,depthwise separable convolutions(DS Convs)are employed to improve convolutional operations’efficiency.Furthermore,several weighted multi-scale residual blocks(WMRBs)are stacked to enhance the multi-scale representation capability.In the reconstruction subnetwork,a group of Conv layers are introduced to filter feature maps to reconstruct the final high-quality image.Extensive experiments were conducted to evaluate the proposed model,and the comparative results with several state-of-the-art algorithms demonstrate the effectiveness of WMRN.展开更多
Unauthorized operations referred to as“black flights”of unmanned aerial vehicles(UAVs)pose a significant danger to public safety,and existing low-attitude object detection algorithms encounter difficulties in balanc...Unauthorized operations referred to as“black flights”of unmanned aerial vehicles(UAVs)pose a significant danger to public safety,and existing low-attitude object detection algorithms encounter difficulties in balancing detection precision and speed.Additionally,their accuracy is insufficient,particularly for small objects in complex environments.To solve these problems,we propose a lightweight feature-enhanced convolutional neural network able to perform detection with high precision detection for low-attitude flying objects in real time to provide guidance information to suppress black-flying UAVs.The proposed network consists of three modules.A lightweight and stable feature extraction module is used to reduce the computational load and stably extract more low-level feature,an enhanced feature processing module significantly improves the feature extraction ability of the model,and an accurate detection module integrates low-level and advanced features to improve the multiscale detection accuracy in complex environments,particularly for small objects.The proposed method achieves a detection speed of 147 frames per second(FPS)and a mean average precision(mAP)of 90.97%for a dataset composed of flying objects,indicating its potential for low-altitude object detection.Furthermore,evaluation results based on microsoft common objects in context(MS COCO)indicate that the proposed method is also applicable to object detection in general.展开更多
Accurately identifying defect patterns in wafer maps can help engineers find abnormal failure factors in production lines.During the wafer testing stage,deep learning methods are widely used in wafer defect detection ...Accurately identifying defect patterns in wafer maps can help engineers find abnormal failure factors in production lines.During the wafer testing stage,deep learning methods are widely used in wafer defect detection due to their powerful feature extraction capa-bilities.However,most of the current wafer defect patterns classification models have high complexity and slow detection speed,which are difficult to apply in the actual wafer production process.In addition,there is a data imbalance in the wafer dataset that seriously affects the training results of the model.To reduce the complexity of the deep model without affecting the wafer feature expression,this paper adjusts the structure of the dense block in the PeleeNet network and proposes a lightweight network WM‐PeleeNet based on the PeleeNet module.In addition,to reduce the impact of data imbalance on model training,this paper proposes a wafer data augmentation method based on a convolutional autoencoder by adding random Gaussian noise to the hidden layer.The method proposed in this paper has an average accuracy of 95.4%on the WM‐811K wafer dataset with only 173.643 KB of the parameters and 316.194 M of FLOPs,and takes only 22.99 s to detect 1000 wafer pictures.Compared with the original PeleeNet network without optimization,the number of parameters and FLOPs are reduced by 92.68%and 58.85%,respectively.Data augmentation on the minority class wafer map improves the average classification accuracy by 1.8%on the WM‐811K dataset.At the same time,the recognition accuracy of minority classes such as Scratch pattern and Donut pattern are significantly improved.展开更多
To solve the problem of difficulty in identifying apple diseases in the natural environment and the low application rate of deep learning recognition networks,a lightweight ResNet(LW-ResNet)model for apple disease rec...To solve the problem of difficulty in identifying apple diseases in the natural environment and the low application rate of deep learning recognition networks,a lightweight ResNet(LW-ResNet)model for apple disease recognition is proposed.Based on the deep residual network(ResNet18),the multi-scale feature extraction layer is constructed by group convolution to realize the compression model and improve the extraction ability of different sizes of lesion features.By improving the identity mapping structure to reduce information loss.By introducing the efficient channel attention module(ECANet)to suppress noise from a complex background.The experimental results show that the average precision,recall and F1-score of the LW-ResNet on the test set are 97.80%,97.92%and 97.85%,respectively.The parameter memory is 2.32 MB,which is 94%less than that of ResNet18.Compared with the classic lightweight networks SqueezeNet and MobileNetV2,LW-ResNet has obvious advantages in recognition performance,speed,parameter memory requirement and time complexity.The proposed model has the advantages of low computational cost,low storage cost,strong real-time performance,high identification accuracy,and strong practicability,which can meet the needs of real-time identification task of apple leaf disease on resource-constrained devices.展开更多
Weather phenomenon recognition plays an important role in the field of meteorology.Nowadays,weather radars and weathers sensor have been widely used for weather recognition.However,given the high cost in deploying and...Weather phenomenon recognition plays an important role in the field of meteorology.Nowadays,weather radars and weathers sensor have been widely used for weather recognition.However,given the high cost in deploying and maintaining the devices,it is difficult to apply them to intensive weather phenomenon recognition.Moreover,advanced machine learning models such as Convolutional Neural Networks(CNNs)have shown a lot of promise in meteorology,but these models also require intensive computation and large memory,which make it difficult to use them in reality.In practice,lightweight models are often used to solve such problems.However,lightweight models often result in significant performance losses.To this end,after taking a deep dive into a large number of lightweight models and summarizing their shortcomings,we propose a novel lightweight CNNs model which is constructed based on new building blocks.The experimental results show that the model proposed in this paper has comparable performance with the mainstream non-lightweight model while also saving 25 times of memory consumption.Such memory reduction is even better than that of existing lightweight models.展开更多
文摘Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.
基金the National Natural Science Foundation of China(Grant Number 62076246).
文摘Human pose estimation aims to localize the body joints from image or video data.With the development of deeplearning,pose estimation has become a hot research topic in the field of computer vision.In recent years,humanpose estimation has achieved great success in multiple fields such as animation and sports.However,to obtainaccurate positioning results,existing methods may suffer from large model sizes,a high number of parameters,and increased complexity,leading to high computing costs.In this paper,we propose a new lightweight featureencoder to construct a high-resolution network that reduces the number of parameters and lowers the computingcost.We also introduced a semantic enhancement module that improves global feature extraction and networkperformance by combining channel and spatial dimensions.Furthermore,we propose a dense connected spatialpyramid pooling module to compensate for the decrease in image resolution and information loss in the network.Finally,ourmethod effectively reduces the number of parameters and complexitywhile ensuring high performance.Extensive experiments show that our method achieves a competitive performance while dramatically reducing thenumber of parameters,and operational complexity.Specifically,our method can obtain 89.9%AP score on MPIIVAL,while the number of parameters and the complexity of operations were reduced by 41%and 36%,respectively.
基金Guangdong Science and Technology Program under Grant No.202206010052Foshan Province R&D Key Project under Grant No.2020001006827Guangdong Academy of Sciences Integrated Industry Technology Innovation Center Action Special Project under Grant No.2022GDASZH-2022010108.
文摘The employment of deep convolutional neural networks has recently contributed to significant progress in single image super-resolution(SISR)research.However,the high computational demands of most SR techniques hinder their applicability to edge devices,despite their satisfactory reconstruction performance.These methods commonly use standard convolutions,which increase the convolutional operation cost of the model.In this paper,a lightweight Partial Separation and Multiscale Fusion Network(PSMFNet)is proposed to alleviate this problem.Specifically,this paper introduces partial convolution(PConv),which reduces the redundant convolution operations throughout the model by separating some of the features of an image while retaining features useful for image reconstruction.Additionally,it is worth noting that the existing methods have not fully utilized the rich feature information,leading to information loss,which reduces the ability to learn feature representations.Inspired by self-attention,this paper develops a multiscale feature fusion block(MFFB),which can better utilize the non-local features of an image.MFFB can learn long-range dependencies from the spatial dimension and extract features from the channel dimension,thereby obtaining more comprehensive and rich feature information.As the role of the MFFB is to capture rich global features,this paper further introduces an efficient inverted residual block(EIRB)to supplement the local feature extraction ability of PSMFNet.A comprehensive analysis of the experimental results shows that PSMFNet maintains a better performance with fewer parameters than the state-of-the-art models.
基金This work was supported by Science and Technology Cooperation Special Project of Shijiazhuang(SJZZXA23005).
文摘In minimally invasive surgery,endoscopes or laparoscopes equipped with miniature cameras and tools are used to enter the human body for therapeutic purposes through small incisions or natural cavities.However,in clinical operating environments,endoscopic images often suffer from challenges such as low texture,uneven illumination,and non-rigid structures,which affect feature observation and extraction.This can severely impact surgical navigation or clinical diagnosis due to missing feature points in endoscopic images,leading to treatment and postoperative recovery issues for patients.To address these challenges,this paper introduces,for the first time,a Cross-Channel Multi-Modal Adaptive Spatial Feature Fusion(ASFF)module based on the lightweight architecture of EfficientViT.Additionally,a novel lightweight feature extraction and matching network based on attention mechanism is proposed.This network dynamically adjusts attention weights for cross-modal information from grayscale images and optical flow images through a dual-branch Siamese network.It extracts static and dynamic information features ranging from low-level to high-level,and from local to global,ensuring robust feature extraction across different widths,noise levels,and blur scenarios.Global and local matching are performed through a multi-level cascaded attention mechanism,with cross-channel attention introduced to simultaneously extract low-level and high-level features.Extensive ablation experiments and comparative studies are conducted on the HyperKvasir,EAD,M2caiSeg,CVC-ClinicDB,and UCL synthetic datasets.Experimental results demonstrate that the proposed network improves upon the baseline EfficientViT-B3 model by 75.4%in accuracy(Acc),while also enhancing runtime performance and storage efficiency.When compared with the complex DenseDescriptor feature extraction network,the difference in Acc is less than 7.22%,and IoU calculation results on specific datasets outperform complex dense models.Furthermore,this method increases the F1 score by 33.2%and accelerates runtime by 70.2%.It is noteworthy that the speed of CMMCAN surpasses that of comparative lightweight models,with feature extraction and matching performance comparable to existing complex models but with faster speed and higher cost-effectiveness.
基金the Liaoning Provincial Department of Education 2021 Annual Scientific Research Funding Program(Grant Numbers LJKZ0535,LJKZ0526)the 2021 Annual Comprehensive Reform of Undergraduate Education Teaching(Grant Numbers JGLX2021020,JCLX2021008)Graduate Innovation Fund of Dalian Polytechnic University(Grant Number 2023CXYJ13).
文摘In pursuit of cost-effective manufacturing,enterprises are increasingly adopting the practice of utilizing recycled semiconductor chips.To ensure consistent chip orientation during packaging,a circular marker on the front side is employed for pin alignment following successful functional testing.However,recycled chips often exhibit substantial surface wear,and the identification of the relatively small marker proves challenging.Moreover,the complexity of generic target detection algorithms hampers seamless deployment.Addressing these issues,this paper introduces a lightweight YOLOv8s-based network tailored for detecting markings on recycled chips,termed Van-YOLOv8.Initially,to alleviate the influence of diminutive,low-resolution markings on the precision of deep learning models,we utilize an upscaling approach for enhanced resolution.This technique relies on the Super-Resolution Generative Adversarial Network with Extended Training(SRGANext)network,facilitating the reconstruction of high-fidelity images that align with input specifications.Subsequently,we replace the original YOLOv8smodel’s backbone feature extraction network with the lightweight VanillaNetwork(VanillaNet),simplifying the branch structure to reduce network parameters.Finally,a Hybrid Attention Mechanism(HAM)is implemented to capture essential details from input images,improving feature representation while concurrently expediting model inference speed.Experimental results demonstrate that the Van-YOLOv8 network outperforms the original YOLOv8s on a recycled chip dataset in various aspects.Significantly,it demonstrates superiority in parameter count,computational intricacy,precision in identifying targets,and speed when compared to certain prevalent algorithms in the current landscape.The proposed approach proves promising for real-time detection of recycled chips in practical factory settings.
基金supported by the National Natural Science Foundation of China(No.62176034)the Science and Technology Research Program of Chongqing Municipal Education Commission(No.KJZD-M202300604)the Natural Science Foundation of Chongqing(Nos.cstc2021jcyj-msxmX0518,2023NSCQ-MSX1781).
文摘Automatic crack detection of cement pavement chiefly benefits from the rapid development of deep learning,with convolutional neural networks(CNN)playing an important role in this field.However,as the performance of crack detection in cement pavement improves,the depth and width of the network structure are significantly increased,which necessitates more computing power and storage space.This limitation hampers the practical implementation of crack detection models on various platforms,particularly portable devices like small mobile devices.To solve these problems,we propose a dual-encoder-based network architecture that focuses on extracting more comprehensive fracture feature information and combines cross-fusion modules and coordinated attention mechanisms formore efficient feature fusion.Firstly,we use small channel convolution to construct shallow feature extractionmodule(SFEM)to extract low-level feature information of cracks in cement pavement images,in order to obtainmore information about cracks in the shallowfeatures of images.In addition,we construct large kernel atrous convolution(LKAC)to enhance crack information,which incorporates coordination attention mechanism for non-crack information filtering,and large kernel atrous convolution with different cores,using different receptive fields to extract more detailed edge and context information.Finally,the three-stage feature map outputs from the shallow feature extraction module is cross-fused with the two-stage feature map outputs from the large kernel atrous convolution module,and the shallow feature and detailed edge feature are fully fused to obtain the final crack prediction map.We evaluate our method on three public crack datasets:DeepCrack,CFD,and Crack500.Experimental results on theDeepCrack dataset demonstrate the effectiveness of our proposed method compared to state-of-the-art crack detection methods,which achieves Precision(P)87.2%,Recall(R)87.7%,and F-score(F1)87.4%.Thanks to our lightweight crack detectionmodel,the parameter count of the model in real-world detection scenarios has been significantly reduced to less than 2M.This advancement also facilitates technical support for portable scene detection.
基金supported by the Shanghai Sailing Program,China (No.20YF1447600)the Research Start-Up Project of Shanghai Institute of Technology (No.YJ2021-60)+1 种基金the Collaborative Innovation Project of Shanghai Institute of Technology (No.XTCX2020-12)the Science and Technology Talent Development Fund for Young and Middle-Aged Teachers at Shanghai Institute of Technology (No.ZQ2022-6)。
文摘There is a problem of real-time detection difficulty in road surface damage detection. This paper proposes an improved lightweight model based on you only look once version 5(YOLOv5). Firstly, this paper fully utilized the convolutional neural network(CNN) + ghosting bottleneck(G_bneck) architecture to reduce redundant feature maps. Afterwards, we upgraded the original upsampling algorithm to content-aware reassembly of features(CARAFE) and increased the receptive field. Finally, we replaced the spatial pyramid pooling fast(SPPF) module with the basic receptive field block(Basic RFB) pooling module and added dilated convolution. After comparative experiments, we can see that the number of parameters and model size of the improved algorithm in this paper have been reduced by nearly half compared to the YOLOv5s. The frame rate per second(FPS) has been increased by 3.25 times. The mean average precision(m AP@0.5: 0.95) has increased by 8%—17% compared to other lightweight algorithms.
基金supported by the National Natural Science Foundation of China(Nos.62373215,62373219 and 62073193)the Natural Science Foundation of Shandong Province(No.ZR2023MF100)+1 种基金the Key Projects of the Ministry of Industry and Information Technology(No.TC220H057-2022)the Independently Developed Instrument Funds of Shandong University(No.zy20240201)。
文摘Current you only look once(YOLO)-based algorithm model is facing the challenge of overwhelming parameters and calculation complexity under the printed circuit board(PCB)defect detection application scenario.In order to solve this problem,we propose a new method,which combined the lightweight network mobile vision transformer(Mobile Vi T)with the convolutional block attention module(CBAM)mechanism and the new regression loss function.This method needed less computation resources,making it more suitable for embedded edge detection devices.Meanwhile,the new loss function improved the positioning accuracy of the bounding box and enhanced the robustness of the model.In addition,experiments on public datasets demonstrate that the improved model achieves an average accuracy of 87.9%across six typical defect detection tasks,while reducing computational costs by nearly 90%.It significantly reduces the model's computational requirements while maintaining accuracy,ensuring reliable performance for edge deployment.
基金supported by the National Natural Science Foundation of China(42202175).
文摘Mineral identification is foundational to geological survey research,mineral resource exploration,and mining engineering.Considering the diversity of mineral types and the challenge of achieving high recognition accuracy for similar features,this study introduces a mineral detection method based on YOLOv8-SBI.This work enhances feature extraction by integrating spatial pyramid pooling-fast(SPPF)with the simplified self-attention module(SimAM),significantly improving the precision of mineral feature detection.In the feature fusion network,a weighted bidirectional feature pyramid network is employed for advanced cross-channel feature integration,effectively reducing feature redundancy.Additionally,Inner-Intersection Over Union(InnerIOU)is used as the loss function to improve the average quality localization performance of anchor boxes.Experimental results show that the YOLOv8-SBI model achieves an accuracy of 67.9%,a recall of 74.3%,a mAP@0.5 of 75.8%,and a mAP@0.5:0.95 of 56.7%,with a real-time detection speed of 244.2 frames per second.Compared to YOLOv8,YOLOv8-SBI demonstrates a significant improvement with 15.4%increase in accuracy,28.5%increase in recall,and increases of 28.1%and 20.9%in mAP@0.5 and mAP@0.5:0.95,respectively.Furthermore,relative to other models,such as YOLOv3,YOLOv5,YOLOv6,YOLOv8,YOLOv9,and YOLOv10,YOLOv8-SBI has a smaller parameter size of only 3.01×10^(6).This highlights the optimal balance between detection accuracy and speed,thereby offering robust technical support for intelligent mineral classification.
基金suported by the Basic Science Research Program through the National Research Foundation of Korea(NRF)funded by the Ministry of Science and ICT(NRF-2022R1A2C2012243).
文摘Accurately identifying crop pests and diseases ensures agricultural productivity and safety.Although current YOLO-based detection models offer real-time capabilities,their conventional convolutional layers involve high computational redundancy and a fixed receptive field,making it challenging to capture local details and global semantics in complex scenarios simultaneously.This leads to significant issues like missed detections of small targets and heightened sensitivity to background interference.To address these challenges,this paper proposes a lightweight adaptive detection network—StarSpark-AdaptiveNet(SSANet),which optimizes features through a dual-module collaborative mechanism.Specifically,the StarNet module utilizes Depthwise separable convolutions(DW-Conv)and dynamic star operations to establish multi-stage feature extraction pathways,enhancing local detail perception within a lightweight framework.Moreover,the Multi-scale Adaptive Spatial Attention Gate(MASAG)module integrates cross-layer feature fusion and dynamic weight allocation to capture multi-scale global contextual information,effectively suppressing background noise.These modules jointly form a“local enhancement-global calibration”bidirectional optimization mechanism,significantly improving the model’s adaptability to complex disease patterns.Furthermore,the proposed Scale-based Dynamic Loss(SD Loss)dynamically adjusts the weight of scale and localization losses,improving regression stability and localization accuracy,especially for small targets.Experiments on the eggplant fruit disease dataset demonstrate that SSANet achieves an mAP50 of 83.9%and a detection speed of 273.5 FPS with only 2.11 M parameters and 5.1 GFLOPs computational cost,outperforming the baseline YOLO11 model by reducing parameters by 18.1%,increasing mAP50 by 1.3%,and improving inference speed by 9.1%.Ablation studies further confirm the effectiveness and complementarity of the modules.SSANet offers a high-accuracy,low-cost solution suitable for real-time pest and disease detection in crops,facilitating edge device deployment and promoting precision agriculture.
文摘This paper introduces a lightweight remote sensing image dehazing network called multidimensional weight regulation network(MDWR-Net), which addresses the high computational cost of existing methods. Previous works, often based on the encoder-decoder structure and utilizing multiple upsampling and downsampling layers, are computationally expensive. To improve efficiency, the paper proposes two modules: the efficient spatial resolution recovery module(ESRR) for upsampling and the efficient depth information augmentation module(EDIA) for downsampling.These modules not only reduce model complexity but also enhance performance. Additionally, the partial feature weight learning module(PFWL) is introduced to reduce the computational burden by applying weight learning across partial dimensions, rather than using full-channel convolution.To overcome the limitations of convolutional neural networks(CNN)-based networks, the haze distribution index transformer(HDIT) is integrated into the decoder. We also propose the physicalbased non-adjacent feature fusion module(PNFF), which leverages the atmospheric scattering model to improve generalization of our MDWR-Net. The MDWR-Net achieves superior dehazing performance with a computational cost of just 2.98×10^(9) multiply-accumulate operations(MACs),which is less than one-tenth of previous methods. Experimental results validate its effectiveness in balancing performance and computational efficiency.
基金supported by the Science and Technology Bureau of Xi’an project(24KGDW0049)the Key Research and Development Programof Shaanxi(2023-YBGY-264)the Key Research and Development Program of Guangxi(GK-AB20159032).
文摘In recent years,the country has spent significant workforce and material resources to prevent traffic accidents,particularly those caused by fatigued driving.The current studies mainly concentrate on driver physiological signals,driving behavior,and vehicle information.However,most of the approaches are computationally intensive and inconvenient for real-time detection.Therefore,this paper designs a network that combines precision,speed and lightweight and proposes an algorithm for facial fatigue detection based on multi-feature fusion.Specifically,the face detection model takes YOLOv8(You Only Look Once version 8)as the basic framework,and replaces its backbone network with MobileNetv3.To focus on the significant regions in the image,CPCA(Channel Prior Convolution Attention)is adopted to enhance the network’s capacity for feature extraction.Meanwhile,the network training phase employs the Focal-EIOU(Focal and Efficient Intersection Over Union)loss function,which makes the network lightweight and increases the accuracy of target detection.Ultimately,the Dlib toolkit was employed to annotate 68 facial feature points.This study established an evaluation metric for facial fatigue and developed a novel fatigue detection algorithm to assess the driver’s condition.A series of comparative experiments were carried out on the self-built dataset.The suggested method’s mAP(mean Average Precision)values for object detection and fatigue detection are 96.71%and 95.75%,respectively,as well as the detection speed is 47 FPS(Frames Per Second).This method can balance the contradiction between computational complexity and model accuracy.Furthermore,it can be transplanted to NVIDIA Jetson Orin NX and quickly detect the driver’s state while maintaining a high degree of accuracy.It contributes to the development of automobile safety systems and reduces the occurrence of traffic accidents.
基金funded by Innovation and Development Special Project of China Meteorological Administration(CXFZ2022J038,CXFZ2024J035)Sichuan Science and Technology Program(No.2023YFQ0072)+1 种基金Key Laboratory of Smart Earth(No.KF2023YB03-07)Automatic Software Generation and Intelligent Service Key Laboratory of Sichuan Province(CUIT-SAG202210).
文摘Accurate cloud classification plays a crucial role in aviation safety,climate monitoring,and localized weather forecasting.Current research has been focusing on machine learning techniques,particularly deep learning based model,for the types identification.However,traditional approaches such as convolutional neural networks(CNNs)encounter difficulties in capturing global contextual information.In addition,they are computationally expensive,which restricts their usability in resource-limited environments.To tackle these issues,we present the Cloud Vision Transformer(CloudViT),a lightweight model that integrates CNNs with Transformers.The integration enables an effective balance between local and global feature extraction.To be specific,CloudViT comprises two innovative modules:Feature Extraction(E_Module)and Downsampling(D_Module).These modules are able to significantly reduce the number of model parameters and computational complexity while maintaining translation invariance and enhancing contextual comprehension.Overall,the CloudViT includes 0.93×10^(6)parameters,which decreases more than ten times compared to the SOTA(State-of-the-Art)model CloudNet.Comprehensive evaluations conducted on the HBMCD and SWIMCAT datasets showcase the outstanding performance of CloudViT.It achieves classification accuracies of 98.45%and 100%,respectively.Moreover,the efficiency and scalability of CloudViT make it an ideal candidate for deployment inmobile cloud observation systems,enabling real-time cloud image classification.The proposed hybrid architecture of CloudViT offers a promising approach for advancing ground-based cloud image classification.It holds significant potential for both optimizing performance and facilitating practical deployment scenarios.
基金funded by the National Natural Science Foundation of China(grant number:62172292).
文摘Vehicle re-identification involves matching images of vehicles across varying camera views.The diversity of camera locations along different roadways leads to significant intra-class variation and only minimal inter-class similarity in the collected vehicle images,which increases the complexity of re-identification tasks.To tackle these challenges,this study proposes AG-GCN(Attention-Guided Graph Convolutional Network),a novel framework integrating several pivotal components.Initially,AG-GCN embeds a lightweight attention module within the ResNet-50 structure to learn feature weights automatically,thereby improving the representation of vehicle features globally by highlighting salient features and suppressing extraneous ones.Moreover,AG-GCN adopts a graph-based structure to encapsulate deep local features.A graph convolutional network then amalgamates these features to understand the relationships among vehicle-related characteristics.Subsequently,we amalgamate feature maps from both the attention and graph-based branches for a more comprehensive representation of vehicle features.The framework then gauges feature similarities and ranks them,thus enhancing the accuracy of vehicle re-identification.Comprehensive qualitative and quantitative analyses on two publicly available datasets verify the efficacy of AG-GCN in addressing intra-class and inter-class variability issues.
基金the National Natural Science Foundation of China(61772149,61866009,61762028,U1701267,61702169)Guangxi Science and Technology Project(2019GXNSFFA245014,ZY20198016,AD18281079,AD18216004)+1 种基金the Natural Science Foundation of Hunan Province(2020JJ3014)Guangxi Colleges and Universities Key Laboratory of Intelligent Processing of Computer Images and Graphics(GIIP202001).
文摘The tradeoff between efficiency and model size of the convolutional neural network(CNN)is an essential issue for applications of CNN-based algorithms to diverse real-world tasks.Although deep learning-based methods have achieved significant improvements in image super-resolution(SR),current CNNbased techniques mainly contain massive parameters and a high computational complexity,limiting their practical applications.In this paper,we present a fast and lightweight framework,named weighted multi-scale residual network(WMRN),for a better tradeoff between SR performance and computational efficiency.With the modified residual structure,depthwise separable convolutions(DS Convs)are employed to improve convolutional operations’efficiency.Furthermore,several weighted multi-scale residual blocks(WMRBs)are stacked to enhance the multi-scale representation capability.In the reconstruction subnetwork,a group of Conv layers are introduced to filter feature maps to reconstruct the final high-quality image.Extensive experiments were conducted to evaluate the proposed model,and the comparative results with several state-of-the-art algorithms demonstrate the effectiveness of WMRN.
基金supported by the National Natural Science Foundation of China(52075027)the Fundamental Research Funds for the Central Universities(2020XJJD03).
文摘Unauthorized operations referred to as“black flights”of unmanned aerial vehicles(UAVs)pose a significant danger to public safety,and existing low-attitude object detection algorithms encounter difficulties in balancing detection precision and speed.Additionally,their accuracy is insufficient,particularly for small objects in complex environments.To solve these problems,we propose a lightweight feature-enhanced convolutional neural network able to perform detection with high precision detection for low-attitude flying objects in real time to provide guidance information to suppress black-flying UAVs.The proposed network consists of three modules.A lightweight and stable feature extraction module is used to reduce the computational load and stably extract more low-level feature,an enhanced feature processing module significantly improves the feature extraction ability of the model,and an accurate detection module integrates low-level and advanced features to improve the multiscale detection accuracy in complex environments,particularly for small objects.The proposed method achieves a detection speed of 147 frames per second(FPS)and a mean average precision(mAP)of 90.97%for a dataset composed of flying objects,indicating its potential for low-altitude object detection.Furthermore,evaluation results based on microsoft common objects in context(MS COCO)indicate that the proposed method is also applicable to object detection in general.
基金supported by a project jointly funded by the Beijing Municipal Education Commission and Municipal Natural Science Foundation under grant KZ202010005004.
文摘Accurately identifying defect patterns in wafer maps can help engineers find abnormal failure factors in production lines.During the wafer testing stage,deep learning methods are widely used in wafer defect detection due to their powerful feature extraction capa-bilities.However,most of the current wafer defect patterns classification models have high complexity and slow detection speed,which are difficult to apply in the actual wafer production process.In addition,there is a data imbalance in the wafer dataset that seriously affects the training results of the model.To reduce the complexity of the deep model without affecting the wafer feature expression,this paper adjusts the structure of the dense block in the PeleeNet network and proposes a lightweight network WM‐PeleeNet based on the PeleeNet module.In addition,to reduce the impact of data imbalance on model training,this paper proposes a wafer data augmentation method based on a convolutional autoencoder by adding random Gaussian noise to the hidden layer.The method proposed in this paper has an average accuracy of 95.4%on the WM‐811K wafer dataset with only 173.643 KB of the parameters and 316.194 M of FLOPs,and takes only 22.99 s to detect 1000 wafer pictures.Compared with the original PeleeNet network without optimization,the number of parameters and FLOPs are reduced by 92.68%and 58.85%,respectively.Data augmentation on the minority class wafer map improves the average classification accuracy by 1.8%on the WM‐811K dataset.At the same time,the recognition accuracy of minority classes such as Scratch pattern and Donut pattern are significantly improved.
基金funded by the Science and Technology Development Program of Jilin Province(20190301024NY)the Precision Agriculture and Big Data Engineering Research Center of Jilin Province(2020C005).
文摘To solve the problem of difficulty in identifying apple diseases in the natural environment and the low application rate of deep learning recognition networks,a lightweight ResNet(LW-ResNet)model for apple disease recognition is proposed.Based on the deep residual network(ResNet18),the multi-scale feature extraction layer is constructed by group convolution to realize the compression model and improve the extraction ability of different sizes of lesion features.By improving the identity mapping structure to reduce information loss.By introducing the efficient channel attention module(ECANet)to suppress noise from a complex background.The experimental results show that the average precision,recall and F1-score of the LW-ResNet on the test set are 97.80%,97.92%and 97.85%,respectively.The parameter memory is 2.32 MB,which is 94%less than that of ResNet18.Compared with the classic lightweight networks SqueezeNet and MobileNetV2,LW-ResNet has obvious advantages in recognition performance,speed,parameter memory requirement and time complexity.The proposed model has the advantages of low computational cost,low storage cost,strong real-time performance,high identification accuracy,and strong practicability,which can meet the needs of real-time identification task of apple leaf disease on resource-constrained devices.
基金This paper is supported by the following funds:National Key R&D Program of China(2018YFF01010100)National natural science foundation of China(61672064)+1 种基金Basic Research Program of Qinghai Province under Grants No.2020-ZJ-709Advanced information network Beijing laboratory(PXM2019_014204_500029).
文摘Weather phenomenon recognition plays an important role in the field of meteorology.Nowadays,weather radars and weathers sensor have been widely used for weather recognition.However,given the high cost in deploying and maintaining the devices,it is difficult to apply them to intensive weather phenomenon recognition.Moreover,advanced machine learning models such as Convolutional Neural Networks(CNNs)have shown a lot of promise in meteorology,but these models also require intensive computation and large memory,which make it difficult to use them in reality.In practice,lightweight models are often used to solve such problems.However,lightweight models often result in significant performance losses.To this end,after taking a deep dive into a large number of lightweight models and summarizing their shortcomings,we propose a novel lightweight CNNs model which is constructed based on new building blocks.The experimental results show that the model proposed in this paper has comparable performance with the mainstream non-lightweight model while also saving 25 times of memory consumption.Such memory reduction is even better than that of existing lightweight models.
