To address critical challenges in nighttime ship detection—high small-target missed detection(over 20%),insufficient lightweighting,and limited generalization due to scarce,low-quality datasets—this study proposes a...To address critical challenges in nighttime ship detection—high small-target missed detection(over 20%),insufficient lightweighting,and limited generalization due to scarce,low-quality datasets—this study proposes a systematic solution.First,a high-quality Night-Ships dataset is constructed via CycleGAN-based day-night transfer,combined with a dual-threshold cleaning strategy(Laplacian variance sharpness filtering and brightness-color deviation screening).Second,a Cross-stage Lightweight Fusion-You Only Look Once version 8(CLF-YOLOv8)is proposed with key improvements:the Neck network is reconstructed by replacing Cross Stage Partial(CSP)structure with the Cross Stage Partial Multi-Scale Convolutional Block(CSP-MSCB)and integrating Bidirectional Feature Pyramid Network(BiFPN)for weighted multi-scale fusion to enhance small-target detection;a Lightweight Shared Convolutional and Separated Batch Normalization Detection-Head(LSCSBD-Head)with shared convolutions and layer-wise Batch Normalization(BN)reduces parameters to 1.8M(42% fewer than YOLOv8n);and the FocalMinimum Point Distance Intersection over Union(Focal-MPDIoU)loss combines Minimum Point Distance Intersection over Union(MPDIoU)geometric constraints and Focal weighting to optimize low-overlap targets.Experiments show CLFYOLOv8 achieves 97.6%mAP@0.5(0.7% higher than YOLOv8n)with 1.8 M parameters,outperforming mainstream models in small-target detection,overlapping target discrimination,and adaptability to complex lighting.展开更多
With the large-scale deployment of the Internet ofThings(IoT)devices,their weak securitymechanisms make them prime targets for malware attacks.Attackers often use Domain Generation Algorithm(DGA)to generate random dom...With the large-scale deployment of the Internet ofThings(IoT)devices,their weak securitymechanisms make them prime targets for malware attacks.Attackers often use Domain Generation Algorithm(DGA)to generate random domain names,hiding the real IP of Command and Control(C&C)servers to build botnets.Due to the randomness and dynamics of DGA,traditional methods struggle to detect them accurately,increasing the difficulty of network defense.This paper proposes a lightweight DGA detection model based on knowledge distillation for resource-constrained IoT environments.Specifically,a teacher model combining CharacterBERT,a bidirectional long short-term memory(BiLSTM)network,and attention mechanism(ATT)is constructed:it extracts character-level semantic features viaCharacterBERT,captures sequence dependencieswith the BiLSTM,and integrates theATT for key feature weighting,formingmulti-granularity feature fusion.An improved knowledge distillation approach transfers the teacher model’s learned knowledge to the simplified DistilBERT student model.Experimental results show the teacher model achieves 98.68%detection accuracy.The student modelmaintains slightly improved accuracy while significantly compressing parameters to approximately 38.4%of the teacher model’s scale,greatly reducing computational overhead for IoT deployment.展开更多
Bird's-eye-view(BEV)perception is a core technology for autonomous driving systems.However,existing solutions face the dilemma of high costs associated with multimodal methods and limited performance of vision-onl...Bird's-eye-view(BEV)perception is a core technology for autonomous driving systems.However,existing solutions face the dilemma of high costs associated with multimodal methods and limited performance of vision-only approaches.To address this issue,this paper proposes a framework named“a lightweight pure visual BEV perception method based on dual distillation of spatial-temporal knowledge”.This framework innovatively designs a lightweight vision-only student model based on Res Net,which leverages a dual distillation mechanism to learn from a powerful teacher model that integrates temporal information from both image and light detection and ranging(LiDAR)modalities.Specifically,we distill efficient multi-modal feature extraction and spatial fusion capabilities from the BEVFusion model,and distill advanced temporal information fusion and spatiotemporal attention mechanisms from the BEVFormer model.This dual distillation strategy enables the student model to achieve perception performance close to that of multi-modal models without relying on Li DAR.Experimental results on the nu Scenes dataset demonstrate that the proposed model significantly outperforms classical vision-only algorithms,achieves comparable performance to current state-of-the-art vision-only methods on the nu Scenes detection leaderboard in terms of both mean average precision(mAP)and the nu Scenes detection score(NDS)metrics,and exhibits notable advantages in inference computational efficiency.Although the proposed dual-teacher paradigm incurs higher offline training costs compared to single-model approaches,it yields a streamlined and highly efficient student model suitable for resource-constrained real-time deployment.This provides an effective pathway toward low-cost,high-performance autonomous driving perception systems.展开更多
Deep learning has been recognized as an effective method for indoor positioning.However,most existing real-valued neural networks(RVNNs)treat the two constituent components of complex-valued channel state information(...Deep learning has been recognized as an effective method for indoor positioning.However,most existing real-valued neural networks(RVNNs)treat the two constituent components of complex-valued channel state information(CSI)as real-valued inputs,potentially discarding useful information embedded in the original CSI.In addition,existing positioning models generally face the contradiction between computational complexity and positioning accuracy.To address these issues,we combine graph neural network(GNN)with complex-valued neural network(CVNN)to construct a lightweight indoor positioning model named CGNet.CGNet employs complexvalued convolution operation to directly process the original CSI data,fully exploiting the correlation between real and imaginary parts of CSI while extracting local features.Subsequently,the feature values are treated as nodes,and conditional position encoding(CPE)module is applied to add positional information.To reduce the number of connections in the graph structure and lower themodel complexity,feature information is mapped to an efficient graph structure through a dynamic axial graph construction(DAGC)method,with global features extracted usingmaximum relative graph convolution(MRConv).Experimental results show that,on the CTW dataset,CGNet achieves a 10%improvement in positioning accuracy compared to existing methods,while the number of model parameters is only 0.8 M.CGNet achieves excellent positioning accuracy with very few parameters.展开更多
Mango is a plant with high economic value in the agricultural industry;thus,it is necessary to maximize the productivity performance of the mango plant,which can be done by implementing artificial intelligence.In this...Mango is a plant with high economic value in the agricultural industry;thus,it is necessary to maximize the productivity performance of the mango plant,which can be done by implementing artificial intelligence.In this study,a lightweight object detection model will be developed that can detect mango plant conditions based on disease potential,so that it becomes an early detection warning system that has an impact on increasing agricultural productivity.The proposed lightweight model integrates YOLOv7-Tiny and the proposed modules,namely the C2S module.The C2S module consists of three sub-modules such as the convolutional block attention module(CBAM),the coordinate attention(CA)module,and the squeeze-and-excitation(SE)module.The dataset is constructed by eight classes,including seven classes of disease conditions and one class of health conditions.The experimental result shows that the proposed lightweight model has the optimal results,which increase by 13.15% of mAP50 compared to the original model YOLOv7-Tiny.While the mAP50:95 also achieved the highest results compared to other models,including YOLOv3-Tiny,YOLOv4-Tiny,YOLOv5,and YOLOv7-Tiny.The advantage of the proposed lightweightmodel is the adaptability that supports it in constrained environments,such as edge computing systems.This proposedmodel can support a robust,precise,and convenient precision agriculture system for the user.展开更多
In printed circuit board(PCB)manufacturing,surface defects can significantly affect product quality.To address the performance degradation,high false detection rates,and missed detections caused by complex backgrounds...In printed circuit board(PCB)manufacturing,surface defects can significantly affect product quality.To address the performance degradation,high false detection rates,and missed detections caused by complex backgrounds in current intelligent inspection algorithms,this paper proposes CG-YOLOv8,a lightweight and improved model based on YOLOv8n for PCB surface defect detection.The proposed method optimizes the network architecture and compresses parameters to reduce model complexity while maintaining high detection accuracy,thereby enhancing the capability of identifying diverse defects under complex conditions.Specifically,a cascaded multi-receptive field(CMRF)module is adopted to replace the SPPF module in the backbone to improve feature perception,and an inverted residual mobile block(IRMB)is integrated into the C2f module to further enhance performance.Additionally,conventional convolution layers are replaced with GSConv to reduce computational cost,and a lightweight Convolutional Block Attention Module based Convolution(CBAMConv)module is introduced after Grouped Spatial Convolution(GSConv)to preserve accuracy through attention mechanisms.The detection head is also optimized by removing medium and large-scale detection layers,thereby enhancing the model’s ability to detect small-scale defects and further reducing complexity.Experimental results show that,compared to the original YOLOv8n,the proposed CG-YOLOv8 reduces parameter count by 53.9%,improves mAP@0.5 by 2.2%,and increases precision and recall by 2.0%and 1.8%,respectively.These improvements demonstrate that CG-YOLOv8 offers an efficient and lightweight solution for PCB surface defect detection.展开更多
As batteries become increasingly essential for energy storage technologies,battery prognosis,and diagnosis remain central to ensure reliable operation and effective management,as well as to aid the in-depth investigat...As batteries become increasingly essential for energy storage technologies,battery prognosis,and diagnosis remain central to ensure reliable operation and effective management,as well as to aid the in-depth investigation of degradation mechanisms.However,dynamic operating conditions,cell-to-cell inconsistencies,and limited availability of labeled data have posed significant challenges to accurate and robust prognosis and diagnosis.Herein,we introduce a time-series-decomposition-based ensembled lightweight learning model(TELL-Me),which employs a synergistic dual-module framework to facilitate accurate and reliable forecasting.The feature module formulates features with physical implications and sheds light on battery aging mechanisms,while the gradient module monitors capacity degradation rates and captures aging trend.TELL-Me achieves high accuracy in end-of-life prediction using minimal historical data from a single battery without requiring offline training dataset,and demonstrates impressive generality and robustness across various operating conditions and battery types.Additionally,by correlating feature contributions with degradation mechanisms across different datasets,TELL-Me is endowed with the diagnostic ability that not only enhances prediction reliability but also provides critical insights into the design and optimization of next-generation batteries.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
Lightweight deep learning models are increasingly required in resource-constrained environments such as mobile devices and the Internet of Medical Things(IoMT).Multi-head convolution with channel attention can facilit...Lightweight deep learning models are increasingly required in resource-constrained environments such as mobile devices and the Internet of Medical Things(IoMT).Multi-head convolution with channel attention can facilitate learning activations relevant to different kernel sizes within a multi-head convolutional layer.Therefore,this study investigates the capability of novel lightweight models incorporating residual multi-head convolution with channel attention(ResMHCNN)blocks to classify medical images.We introduced three novel lightweight deep learning models(BT-Net,LCC-Net,and BC-Net)utilizing the ResMHCNN block as their backbone.These models were crossvalidated and tested on three publicly available medical image datasets:a brain tumor dataset from Figshare consisting of T1-weighted magnetic resonance imaging slices of meningioma,glioma,and pituitary tumors;the LC25000 dataset,which includes microscopic images of lung and colon cancers;and the BreaKHis dataset,containing benign and malignant breast microscopic images.The lightweight models achieved accuracies of 96.9%for 3-class brain tumor classification using BT-Net,and 99.7%for 5-class lung and colon cancer classification using LCC-Net.For 2-class breast cancer classification,BC-Net achieved an accuracy of 96.7%.The parameter counts for the proposed lightweight models—LCC-Net,BC-Net,and BT-Net—are 0.528,0.226,and 1.154 million,respectively.The presented lightweight models,featuring ResMHCNN blocks,may be effectively employed for accurate medical image classification.In the future,these models might be tested for viability in resource-constrained systems such as mobile devices and IoMT platforms.展开更多
Segmenting skin lesions is critical for early skin cancer detection.Existing CNN and Transformer-based methods face challenges such as high computational complexity and limited adaptability to variations in lesion siz...Segmenting skin lesions is critical for early skin cancer detection.Existing CNN and Transformer-based methods face challenges such as high computational complexity and limited adaptability to variations in lesion sizes.To overcome these limitations,we introduce MSAMamba-UNet,a lightweight model that integrates two novel architectures:Multi-Scale Mamba(MSMamba)and Adaptive Dynamic Gating Block(ADGB).MSMamba utilizes multi-scale decomposition and a parallel hierarchical structure to enhance the delineation of irregular lesion boundaries and sensitivity to small targets.ADGB dynamically selects convolutional kernels with varying receptive fields based on input features,improving the model’s capacity to accommodate diverse lesion textures and scales.Additionally,we introduce a Mix Attention Fusion Block(MAF)to enhance shallow feature representation by integrating parallel channel and pixel attention mechanisms.Extensive evaluation of MSAMamba-UNet on the ISIC 2016,ISIC 2017,and ISIC 2018 datasets demonstrates competitive segmentation accuracy with only 0.056 M parameters and 0.069 GFLOPs.Our experiments revealed that MSAMamba-UNet achieved IoU scores of 85.53%,85.47%,and 82.22%,as well as DSC scores of 92.20%,92.17%,and 90.24%,respectively.These results underscore the lightweight design and effectiveness of MSAMamba-UNet.展开更多
The identification of defect types and their reduction values is the most crucial step in coffee grading.In Ethiopia,the current coffee defect investigation techniques rely on manual screening,which requires substanti...The identification of defect types and their reduction values is the most crucial step in coffee grading.In Ethiopia,the current coffee defect investigation techniques rely on manual screening,which requires substantial human resources,time-consuming,and prone to errors.Recently,the deep learning driven object detection has shown promising results in coffee defect identification and grading tasks.In this study,we propose KN-YOLOv8,a modified You Only Look Once version-8(YOLOv8)model optimized for real-time detection of coffee bean defects.This lightweight network incorporates effective feature fusion techniques to accurately detect and locate defects,even among overlapping beans.We have compiled a custom dataset of 562 images comprising thirteen distinct types of defects.The model achieved exceptional performance,with training dataset metrics of 97% recall,100% precision,and 98% mean average precision(mAP).On the test dataset,it maintained outstanding results with 99% recall,100% precision,and 98.9% mAP.The model outperforms existing approaches by achieving a 97.7%m AP for all classes at a 0.5 threshold,while maintaining an optimal precision-recall balance.The model outperforms new approaches by achieving a balance between precision and recall,achieving a mean average precision of 97.7% for all classes.This solution significantly reduces reliance on labor-intensivemanual inspection while improving accuracy.Its lightweight design and high speed make it suitable for real-time industrial applications,transforming coffee quality inspection.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
Efficient three-dimensional(3D)building reconstruction from drone imagery often faces data acquisition,storage,and computational challenges because of its reliance on dense point clouds.In this study,we introduced a n...Efficient three-dimensional(3D)building reconstruction from drone imagery often faces data acquisition,storage,and computational challenges because of its reliance on dense point clouds.In this study,we introduced a novel method for efficient and lightweight 3D building reconstruction from drone imagery using line clouds and sparse point clouds.Our approach eliminates the need to generate dense point clouds,and thus significantly reduces the computational burden by reconstructing 3D models directly from sparse data.We addressed the limitations of line clouds for plane detection and reconstruction by using a new algorithm.This algorithm projects 3D line clouds onto a 2D plane,clusters the projections to identify potential planes,and refines them using sparse point clouds to ensure an accurate and efficient model reconstruction.Extensive qualitative and quantitative experiments demonstrated the effectiveness of our method,demonstrating its superiority over existing techniques in terms of simplicity and efficiency.展开更多
The study presents AW-HRNet,a lightweight high-resolution crack segmentation network that couples Adaptive residual enhancement(AREM)in the spatial domain with Wavelet-based decomposition-reconstruction(WDRM)in the fr...The study presents AW-HRNet,a lightweight high-resolution crack segmentation network that couples Adaptive residual enhancement(AREM)in the spatial domain with Wavelet-based decomposition-reconstruction(WDRM)in the frequency domain.AREM introduces a learnable channel-wise scaling after standard 3×3 convolution and merges it through a residual path to stabilize crack-sensitive responses while suppressing noise.WDRM performs DWT to decouple LL/LH/HL/HH sub-bands,conducts lightweight cross-band fusion,and applies IDWT to restore detail-enhanced features,unifying global topology and boundary sharpness without deformable offsets.Integrated into a high-resolution backbone with auxiliary deep supervision,AW-HRNet attains 79.07%mIoU on CrackSeg9k with only 1.24M parameters and 0.73 GFLOPs,offering an excellent accuracy-efficiency trade-off and strong robustness for real-world deployment.展开更多
The task of student action recognition in the classroom is to precisely capture and analyze the actions of students in classroom videos,providing a foundation for realizing intelligent and accurate teaching.However,th...The task of student action recognition in the classroom is to precisely capture and analyze the actions of students in classroom videos,providing a foundation for realizing intelligent and accurate teaching.However,the complex nature of the classroom environment has added challenges and difficulties in the process of student action recognition.In this research article,with regard to the circumstances where students are prone to be occluded and classroom computing resources are restricted in real classroom scenarios,a lightweight multi-modal fusion action recognition approach is put forward.This proposed method is capable of enhancing the accuracy of student action recognition while concurrently diminishing the number of parameters of the model and the Computation Amount,thereby achieving a more efficient and accurate recognition performance.In the feature extraction stage,this method fuses the keypoint heatmap with the RGB(Red-Green-Blue color model)image.In order to fully utilize the unique information of different modalities for feature complementarity,a Feature Fusion Module(FFE)is introduced.The FFE encodes and fuses the unique features of the two modalities during the feature extraction process.This fusion strategy not only achieves fusion and complementarity between modalities,but also improves the overall model performance.Furthermore,to reduce the computational load and parameter scale of the model,we use keypoint information to crop RGB images.At the same time,the first three networks of the lightweight feature extraction network X3D are used to extract dual-branch features.These methods significantly reduce the computational load and parameter scale.The number of parameters of the model is 1.40 million,and the computation amount is 5.04 billion floating-point operations per second(GFLOPs),achieving an efficient lightweight design.In the Student Classroom Action Dataset(SCAD),the accuracy of the model is 88.36%.In NTU 60(Nanyang Technological University Red-Green-Blue-Depth RGB+Ddataset with 60 categories),the accuracies on X-Sub(The people in the training set are different from those in the test set)and X-View(The perspectives of the training set and the test set are different)are 95.76%and 98.82%,respectively.On the NTU 120 dataset(Nanyang Technological University Red-Green-Blue-Depth dataset with 120 categories),RGB+Dthe accuracies on X-Sub and X-Set(the perspectives of the training set and the test set are different)are 91.97%and 93.45%,respectively.The model has achieved a balance in terms of accuracy,computation amount,and the number of parameters.展开更多
Tomato is a major economic crop worldwide,and diseases on tomato leaves can significantly reduce both yield and quality.Traditional manual inspection is inefficient and highly subjective,making it difficult to meet th...Tomato is a major economic crop worldwide,and diseases on tomato leaves can significantly reduce both yield and quality.Traditional manual inspection is inefficient and highly subjective,making it difficult to meet the requirements of early disease identification in complex natural environments.To address this issue,this study proposes an improved YOLO11-based model,YOLO-SPDNet(Scale Sequence Fusion,Position-Channel Attention,and Dual Enhancement Network).The model integrates the SEAM(Self-Ensembling Attention Mechanism)semantic enhancement module,the MLCA(Mixed Local Channel Attention)lightweight attention mechanism,and the SPA(Scale-Position-Detail Awareness)module composed of SSFF(Scale Sequence Feature Fusion),TFE(Triple Feature Encoding),and CPAM(Channel and Position Attention Mechanism).These enhancements strengthen fine-grained lesion detection while maintaining model lightweightness.Experimental results show that YOLO-SPDNet achieves an accuracy of 91.8%,a recall of 86.5%,and an mAP@0.5 of 90.6%on the test set,with a computational complexity of 12.5 GFLOPs.Furthermore,the model reaches a real-time inference speed of 987 FPS,making it suitable for deployment on mobile agricultural terminals and online monitoring systems.Comparative analysis and ablation studies further validate the reliability and practical applicability of the proposed model in complex natural scenes.展开更多
In this paper,a novel method of ultra-lightweight convolution neural network(CNN)design based on neural architecture search(NAS)and knowledge distillation(KD)is proposed.It can realize the automatic construction of th...In this paper,a novel method of ultra-lightweight convolution neural network(CNN)design based on neural architecture search(NAS)and knowledge distillation(KD)is proposed.It can realize the automatic construction of the space target inverse synthetic aperture radar(ISAR)image recognition model with ultra-lightweight and high accuracy.This method introduces the NAS method into the radar image recognition for the first time,which solves the time-consuming and labor-consuming problems in the artificial design of the space target ISAR image automatic recognition model(STIIARM).On this basis,the NAS model’s knowledge is transferred to the student model with lower computational complexity by the flow of the solution procedure(FSP)distillation method.Thus,the decline of recognition accuracy caused by the direct compression of model structural parameters can be effectively avoided,and the ultralightweight STIIARM can be obtained.In the method,the Inverted Linear Bottleneck(ILB)and Inverted Residual Block(IRB)are firstly taken as each block’s basic structure in CNN.And the expansion ratio,output filter size,number of IRBs,and convolution kernel size are set as the search parameters to construct a hierarchical decomposition search space.Then,the recognition accuracy and computational complexity are taken as the objective function and constraint conditions,respectively,and the global optimization model of the CNN architecture search is established.Next,the simulated annealing(SA)algorithm is used as the search strategy to search out the lightweight and high accuracy STIIARM directly.After that,based on the three principles of similar block structure,the same corresponding channel number,and the minimum computational complexity,the more lightweight student model is designed,and the FSP matrix pairing between the NAS model and student model is completed.Finally,by minimizing the loss between the FSP matrix pairs of the NAS model and student model,the student model’s weight adjustment is completed.Thus the ultra-lightweight and high accuracy STIIARM is obtained.The proposed method’s effectiveness is verified by the simulation experiments on the ISAR image dataset of five types of space targets.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
There is no unified planning standard for unstructured roads,and the morphological structures of these roads are complex and varied.It is important to maintain a balance between accuracy and speed for unstructured roa...There is no unified planning standard for unstructured roads,and the morphological structures of these roads are complex and varied.It is important to maintain a balance between accuracy and speed for unstructured road extraction models.Unstructured road extraction algorithms based on deep learning have problems such as high model complexity,high computational cost,and the inability to adapt to current edge computing devices.Therefore,it is best to use lightweight network models.Considering the need for lightweight models and the characteristics of unstructured roads with different pattern shapes,such as blocks and strips,a TMB(Triple Multi-Block)feature extraction module is proposed,and the overall structure of the TMBNet network is described.The TMB module was compared with SS-nbt,Non-bottleneck-1D,and other modules via experiments.The feasibility and effectiveness of the TMB module design were proven through experiments and visualizations.The comparison experiment,using multiple convolution kernel categories,proved that the TMB module can improve the segmentation accuracy of the network.The comparison with different semantic segmentation networks demonstrates that the TMBNet network has advantages in terms of unstructured road extraction.展开更多
To avoid colliding with trees during its operation,a lawn mower robot must detect the trees.Existing tree detection methods suffer from low detection accuracy(missed detection)and the lack of a lightweight model.In th...To avoid colliding with trees during its operation,a lawn mower robot must detect the trees.Existing tree detection methods suffer from low detection accuracy(missed detection)and the lack of a lightweight model.In this study,a dataset of trees was constructed on the basis of a real lawn environment.According to the theory of channel incremental depthwise convolution and residual suppression,the Embedded-A module is proposed,which expands the depth of the feature map twice to form a residual structure to improve the lightweight degree of the model.According to residual fusion theory,the Embedded-B module is proposed,which improves the accuracy of feature-map downsampling by depthwise convolution and pooling fusion.The Embedded YOLO object detection network is formed by stacking the embedded modules and the fusion of feature maps of different resolutions.Experimental results on the testing set show that the Embedded YOLO tree detection algorithm has 84.17%and 69.91%average precision values respectively for trunk and spherical tree,and 77.04% mean average precision value.The number of convolution parameters is 1.78×10^(6),and the calculation amount is 3.85 billion float operations per second.The size of weight file is 7.11MB,and the detection speed can reach 179 frame/s.This study provides a theoretical basis for the lightweight application of the object detection algorithm based on deep learning for lawn mower robots.展开更多
基金the Shandong Provincial Key Research and Development Program(Grant No.2024SFGC0201).
文摘To address critical challenges in nighttime ship detection—high small-target missed detection(over 20%),insufficient lightweighting,and limited generalization due to scarce,low-quality datasets—this study proposes a systematic solution.First,a high-quality Night-Ships dataset is constructed via CycleGAN-based day-night transfer,combined with a dual-threshold cleaning strategy(Laplacian variance sharpness filtering and brightness-color deviation screening).Second,a Cross-stage Lightweight Fusion-You Only Look Once version 8(CLF-YOLOv8)is proposed with key improvements:the Neck network is reconstructed by replacing Cross Stage Partial(CSP)structure with the Cross Stage Partial Multi-Scale Convolutional Block(CSP-MSCB)and integrating Bidirectional Feature Pyramid Network(BiFPN)for weighted multi-scale fusion to enhance small-target detection;a Lightweight Shared Convolutional and Separated Batch Normalization Detection-Head(LSCSBD-Head)with shared convolutions and layer-wise Batch Normalization(BN)reduces parameters to 1.8M(42% fewer than YOLOv8n);and the FocalMinimum Point Distance Intersection over Union(Focal-MPDIoU)loss combines Minimum Point Distance Intersection over Union(MPDIoU)geometric constraints and Focal weighting to optimize low-overlap targets.Experiments show CLFYOLOv8 achieves 97.6%mAP@0.5(0.7% higher than YOLOv8n)with 1.8 M parameters,outperforming mainstream models in small-target detection,overlapping target discrimination,and adaptability to complex lighting.
基金supported by the following projects:National Natural Science Foundation of China(62461041)Natural Science Foundation of Jiangxi Province China(20242BAB25068).
文摘With the large-scale deployment of the Internet ofThings(IoT)devices,their weak securitymechanisms make them prime targets for malware attacks.Attackers often use Domain Generation Algorithm(DGA)to generate random domain names,hiding the real IP of Command and Control(C&C)servers to build botnets.Due to the randomness and dynamics of DGA,traditional methods struggle to detect them accurately,increasing the difficulty of network defense.This paper proposes a lightweight DGA detection model based on knowledge distillation for resource-constrained IoT environments.Specifically,a teacher model combining CharacterBERT,a bidirectional long short-term memory(BiLSTM)network,and attention mechanism(ATT)is constructed:it extracts character-level semantic features viaCharacterBERT,captures sequence dependencieswith the BiLSTM,and integrates theATT for key feature weighting,formingmulti-granularity feature fusion.An improved knowledge distillation approach transfers the teacher model’s learned knowledge to the simplified DistilBERT student model.Experimental results show the teacher model achieves 98.68%detection accuracy.The student modelmaintains slightly improved accuracy while significantly compressing parameters to approximately 38.4%of the teacher model’s scale,greatly reducing computational overhead for IoT deployment.
基金supported by the National Natural Science Foundation of China(42476084,62203456,42276199)the Stable Support Project of National Key Laboratory(WDZC 20245250302)the National Key R&D Program of China(2024YFC2813502,2024YFC2813302)。
文摘Bird's-eye-view(BEV)perception is a core technology for autonomous driving systems.However,existing solutions face the dilemma of high costs associated with multimodal methods and limited performance of vision-only approaches.To address this issue,this paper proposes a framework named“a lightweight pure visual BEV perception method based on dual distillation of spatial-temporal knowledge”.This framework innovatively designs a lightweight vision-only student model based on Res Net,which leverages a dual distillation mechanism to learn from a powerful teacher model that integrates temporal information from both image and light detection and ranging(LiDAR)modalities.Specifically,we distill efficient multi-modal feature extraction and spatial fusion capabilities from the BEVFusion model,and distill advanced temporal information fusion and spatiotemporal attention mechanisms from the BEVFormer model.This dual distillation strategy enables the student model to achieve perception performance close to that of multi-modal models without relying on Li DAR.Experimental results on the nu Scenes dataset demonstrate that the proposed model significantly outperforms classical vision-only algorithms,achieves comparable performance to current state-of-the-art vision-only methods on the nu Scenes detection leaderboard in terms of both mean average precision(mAP)and the nu Scenes detection score(NDS)metrics,and exhibits notable advantages in inference computational efficiency.Although the proposed dual-teacher paradigm incurs higher offline training costs compared to single-model approaches,it yields a streamlined and highly efficient student model suitable for resource-constrained real-time deployment.This provides an effective pathway toward low-cost,high-performance autonomous driving perception systems.
文摘Deep learning has been recognized as an effective method for indoor positioning.However,most existing real-valued neural networks(RVNNs)treat the two constituent components of complex-valued channel state information(CSI)as real-valued inputs,potentially discarding useful information embedded in the original CSI.In addition,existing positioning models generally face the contradiction between computational complexity and positioning accuracy.To address these issues,we combine graph neural network(GNN)with complex-valued neural network(CVNN)to construct a lightweight indoor positioning model named CGNet.CGNet employs complexvalued convolution operation to directly process the original CSI data,fully exploiting the correlation between real and imaginary parts of CSI while extracting local features.Subsequently,the feature values are treated as nodes,and conditional position encoding(CPE)module is applied to add positional information.To reduce the number of connections in the graph structure and lower themodel complexity,feature information is mapped to an efficient graph structure through a dynamic axial graph construction(DAGC)method,with global features extracted usingmaximum relative graph convolution(MRConv).Experimental results show that,on the CTW dataset,CGNet achieves a 10%improvement in positioning accuracy compared to existing methods,while the number of model parameters is only 0.8 M.CGNet achieves excellent positioning accuracy with very few parameters.
基金supported by National Science and Technology Council(NSTC)Taiwan,Grant No.NSTC 113-2221-E-167-023.
文摘Mango is a plant with high economic value in the agricultural industry;thus,it is necessary to maximize the productivity performance of the mango plant,which can be done by implementing artificial intelligence.In this study,a lightweight object detection model will be developed that can detect mango plant conditions based on disease potential,so that it becomes an early detection warning system that has an impact on increasing agricultural productivity.The proposed lightweight model integrates YOLOv7-Tiny and the proposed modules,namely the C2S module.The C2S module consists of three sub-modules such as the convolutional block attention module(CBAM),the coordinate attention(CA)module,and the squeeze-and-excitation(SE)module.The dataset is constructed by eight classes,including seven classes of disease conditions and one class of health conditions.The experimental result shows that the proposed lightweight model has the optimal results,which increase by 13.15% of mAP50 compared to the original model YOLOv7-Tiny.While the mAP50:95 also achieved the highest results compared to other models,including YOLOv3-Tiny,YOLOv4-Tiny,YOLOv5,and YOLOv7-Tiny.The advantage of the proposed lightweightmodel is the adaptability that supports it in constrained environments,such as edge computing systems.This proposedmodel can support a robust,precise,and convenient precision agriculture system for the user.
基金funded by the Joint Funds of the National Natural Science Foundation of China(U2341223)the Beijing Municipal Natural Science Foundation(No.4232067).
文摘In printed circuit board(PCB)manufacturing,surface defects can significantly affect product quality.To address the performance degradation,high false detection rates,and missed detections caused by complex backgrounds in current intelligent inspection algorithms,this paper proposes CG-YOLOv8,a lightweight and improved model based on YOLOv8n for PCB surface defect detection.The proposed method optimizes the network architecture and compresses parameters to reduce model complexity while maintaining high detection accuracy,thereby enhancing the capability of identifying diverse defects under complex conditions.Specifically,a cascaded multi-receptive field(CMRF)module is adopted to replace the SPPF module in the backbone to improve feature perception,and an inverted residual mobile block(IRMB)is integrated into the C2f module to further enhance performance.Additionally,conventional convolution layers are replaced with GSConv to reduce computational cost,and a lightweight Convolutional Block Attention Module based Convolution(CBAMConv)module is introduced after Grouped Spatial Convolution(GSConv)to preserve accuracy through attention mechanisms.The detection head is also optimized by removing medium and large-scale detection layers,thereby enhancing the model’s ability to detect small-scale defects and further reducing complexity.Experimental results show that,compared to the original YOLOv8n,the proposed CG-YOLOv8 reduces parameter count by 53.9%,improves mAP@0.5 by 2.2%,and increases precision and recall by 2.0%and 1.8%,respectively.These improvements demonstrate that CG-YOLOv8 offers an efficient and lightweight solution for PCB surface defect detection.
基金supported by the National Natural Science Foundation of China(22379021 and 22479021)。
文摘As batteries become increasingly essential for energy storage technologies,battery prognosis,and diagnosis remain central to ensure reliable operation and effective management,as well as to aid the in-depth investigation of degradation mechanisms.However,dynamic operating conditions,cell-to-cell inconsistencies,and limited availability of labeled data have posed significant challenges to accurate and robust prognosis and diagnosis.Herein,we introduce a time-series-decomposition-based ensembled lightweight learning model(TELL-Me),which employs a synergistic dual-module framework to facilitate accurate and reliable forecasting.The feature module formulates features with physical implications and sheds light on battery aging mechanisms,while the gradient module monitors capacity degradation rates and captures aging trend.TELL-Me achieves high accuracy in end-of-life prediction using minimal historical data from a single battery without requiring offline training dataset,and demonstrates impressive generality and robustness across various operating conditions and battery types.Additionally,by correlating feature contributions with degradation mechanisms across different datasets,TELL-Me is endowed with the diagnostic ability that not only enhances prediction reliability but also provides critical insights into the design and optimization of next-generation batteries.
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 Institute of Information&Communications Technology Planning&Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government(MSIT)(IITP-2025-RS-2023-00259678)by INHA UNIVERSITY Research Grant.
文摘Lightweight deep learning models are increasingly required in resource-constrained environments such as mobile devices and the Internet of Medical Things(IoMT).Multi-head convolution with channel attention can facilitate learning activations relevant to different kernel sizes within a multi-head convolutional layer.Therefore,this study investigates the capability of novel lightweight models incorporating residual multi-head convolution with channel attention(ResMHCNN)blocks to classify medical images.We introduced three novel lightweight deep learning models(BT-Net,LCC-Net,and BC-Net)utilizing the ResMHCNN block as their backbone.These models were crossvalidated and tested on three publicly available medical image datasets:a brain tumor dataset from Figshare consisting of T1-weighted magnetic resonance imaging slices of meningioma,glioma,and pituitary tumors;the LC25000 dataset,which includes microscopic images of lung and colon cancers;and the BreaKHis dataset,containing benign and malignant breast microscopic images.The lightweight models achieved accuracies of 96.9%for 3-class brain tumor classification using BT-Net,and 99.7%for 5-class lung and colon cancer classification using LCC-Net.For 2-class breast cancer classification,BC-Net achieved an accuracy of 96.7%.The parameter counts for the proposed lightweight models—LCC-Net,BC-Net,and BT-Net—are 0.528,0.226,and 1.154 million,respectively.The presented lightweight models,featuring ResMHCNN blocks,may be effectively employed for accurate medical image classification.In the future,these models might be tested for viability in resource-constrained systems such as mobile devices and IoMT platforms.
基金supported in part by the National Natural Science Foundation of China under Grant 62201201the Foundation of Henan Educational Committee under Grant 242102211042.
文摘Segmenting skin lesions is critical for early skin cancer detection.Existing CNN and Transformer-based methods face challenges such as high computational complexity and limited adaptability to variations in lesion sizes.To overcome these limitations,we introduce MSAMamba-UNet,a lightweight model that integrates two novel architectures:Multi-Scale Mamba(MSMamba)and Adaptive Dynamic Gating Block(ADGB).MSMamba utilizes multi-scale decomposition and a parallel hierarchical structure to enhance the delineation of irregular lesion boundaries and sensitivity to small targets.ADGB dynamically selects convolutional kernels with varying receptive fields based on input features,improving the model’s capacity to accommodate diverse lesion textures and scales.Additionally,we introduce a Mix Attention Fusion Block(MAF)to enhance shallow feature representation by integrating parallel channel and pixel attention mechanisms.Extensive evaluation of MSAMamba-UNet on the ISIC 2016,ISIC 2017,and ISIC 2018 datasets demonstrates competitive segmentation accuracy with only 0.056 M parameters and 0.069 GFLOPs.Our experiments revealed that MSAMamba-UNet achieved IoU scores of 85.53%,85.47%,and 82.22%,as well as DSC scores of 92.20%,92.17%,and 90.24%,respectively.These results underscore the lightweight design and effectiveness of MSAMamba-UNet.
文摘The identification of defect types and their reduction values is the most crucial step in coffee grading.In Ethiopia,the current coffee defect investigation techniques rely on manual screening,which requires substantial human resources,time-consuming,and prone to errors.Recently,the deep learning driven object detection has shown promising results in coffee defect identification and grading tasks.In this study,we propose KN-YOLOv8,a modified You Only Look Once version-8(YOLOv8)model optimized for real-time detection of coffee bean defects.This lightweight network incorporates effective feature fusion techniques to accurately detect and locate defects,even among overlapping beans.We have compiled a custom dataset of 562 images comprising thirteen distinct types of defects.The model achieved exceptional performance,with training dataset metrics of 97% recall,100% precision,and 98% mean average precision(mAP).On the test dataset,it maintained outstanding results with 99% recall,100% precision,and 98.9% mAP.The model outperforms existing approaches by achieving a 97.7%m AP for all classes at a 0.5 threshold,while maintaining an optimal precision-recall balance.The model outperforms new approaches by achieving a balance between precision and recall,achieving a mean average precision of 97.7% for all classes.This solution significantly reduces reliance on labor-intensivemanual inspection while improving accuracy.Its lightweight design and high speed make it suitable for real-time industrial applications,transforming coffee quality inspection.
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Supported by the Guangdong Major Project of Basic and Applied Basic Research (2023B0303000016)the National Natural Science Foundation of China (U21A20515)。
文摘Efficient three-dimensional(3D)building reconstruction from drone imagery often faces data acquisition,storage,and computational challenges because of its reliance on dense point clouds.In this study,we introduced a novel method for efficient and lightweight 3D building reconstruction from drone imagery using line clouds and sparse point clouds.Our approach eliminates the need to generate dense point clouds,and thus significantly reduces the computational burden by reconstructing 3D models directly from sparse data.We addressed the limitations of line clouds for plane detection and reconstruction by using a new algorithm.This algorithm projects 3D line clouds onto a 2D plane,clusters the projections to identify potential planes,and refines them using sparse point clouds to ensure an accurate and efficient model reconstruction.Extensive qualitative and quantitative experiments demonstrated the effectiveness of our method,demonstrating its superiority over existing techniques in terms of simplicity and efficiency.
文摘The study presents AW-HRNet,a lightweight high-resolution crack segmentation network that couples Adaptive residual enhancement(AREM)in the spatial domain with Wavelet-based decomposition-reconstruction(WDRM)in the frequency domain.AREM introduces a learnable channel-wise scaling after standard 3×3 convolution and merges it through a residual path to stabilize crack-sensitive responses while suppressing noise.WDRM performs DWT to decouple LL/LH/HL/HH sub-bands,conducts lightweight cross-band fusion,and applies IDWT to restore detail-enhanced features,unifying global topology and boundary sharpness without deformable offsets.Integrated into a high-resolution backbone with auxiliary deep supervision,AW-HRNet attains 79.07%mIoU on CrackSeg9k with only 1.24M parameters and 0.73 GFLOPs,offering an excellent accuracy-efficiency trade-off and strong robustness for real-world deployment.
基金supported by the National Natural Science Foundation of China under Grant 62107034the Major Science and Technology Project of Yunnan Province(202402AD080002)Yunnan International Joint R&D Center of China-Laos-Thailand Educational Digitalization(202203AP140006).
文摘The task of student action recognition in the classroom is to precisely capture and analyze the actions of students in classroom videos,providing a foundation for realizing intelligent and accurate teaching.However,the complex nature of the classroom environment has added challenges and difficulties in the process of student action recognition.In this research article,with regard to the circumstances where students are prone to be occluded and classroom computing resources are restricted in real classroom scenarios,a lightweight multi-modal fusion action recognition approach is put forward.This proposed method is capable of enhancing the accuracy of student action recognition while concurrently diminishing the number of parameters of the model and the Computation Amount,thereby achieving a more efficient and accurate recognition performance.In the feature extraction stage,this method fuses the keypoint heatmap with the RGB(Red-Green-Blue color model)image.In order to fully utilize the unique information of different modalities for feature complementarity,a Feature Fusion Module(FFE)is introduced.The FFE encodes and fuses the unique features of the two modalities during the feature extraction process.This fusion strategy not only achieves fusion and complementarity between modalities,but also improves the overall model performance.Furthermore,to reduce the computational load and parameter scale of the model,we use keypoint information to crop RGB images.At the same time,the first three networks of the lightweight feature extraction network X3D are used to extract dual-branch features.These methods significantly reduce the computational load and parameter scale.The number of parameters of the model is 1.40 million,and the computation amount is 5.04 billion floating-point operations per second(GFLOPs),achieving an efficient lightweight design.In the Student Classroom Action Dataset(SCAD),the accuracy of the model is 88.36%.In NTU 60(Nanyang Technological University Red-Green-Blue-Depth RGB+Ddataset with 60 categories),the accuracies on X-Sub(The people in the training set are different from those in the test set)and X-View(The perspectives of the training set and the test set are different)are 95.76%and 98.82%,respectively.On the NTU 120 dataset(Nanyang Technological University Red-Green-Blue-Depth dataset with 120 categories),RGB+Dthe accuracies on X-Sub and X-Set(the perspectives of the training set and the test set are different)are 91.97%and 93.45%,respectively.The model has achieved a balance in terms of accuracy,computation amount,and the number of parameters.
基金Tianmin Tianyuan Boutique Vegetable Industry Technology Service Station(Grant No.2024120011003081)Development of Environmental Monitoring and Traceability System for Wuqing Agricultural Production Areas(Grant No.2024120011001866)。
文摘Tomato is a major economic crop worldwide,and diseases on tomato leaves can significantly reduce both yield and quality.Traditional manual inspection is inefficient and highly subjective,making it difficult to meet the requirements of early disease identification in complex natural environments.To address this issue,this study proposes an improved YOLO11-based model,YOLO-SPDNet(Scale Sequence Fusion,Position-Channel Attention,and Dual Enhancement Network).The model integrates the SEAM(Self-Ensembling Attention Mechanism)semantic enhancement module,the MLCA(Mixed Local Channel Attention)lightweight attention mechanism,and the SPA(Scale-Position-Detail Awareness)module composed of SSFF(Scale Sequence Feature Fusion),TFE(Triple Feature Encoding),and CPAM(Channel and Position Attention Mechanism).These enhancements strengthen fine-grained lesion detection while maintaining model lightweightness.Experimental results show that YOLO-SPDNet achieves an accuracy of 91.8%,a recall of 86.5%,and an mAP@0.5 of 90.6%on the test set,with a computational complexity of 12.5 GFLOPs.Furthermore,the model reaches a real-time inference speed of 987 FPS,making it suitable for deployment on mobile agricultural terminals and online monitoring systems.Comparative analysis and ablation studies further validate the reliability and practical applicability of the proposed model in complex natural scenes.
文摘In this paper,a novel method of ultra-lightweight convolution neural network(CNN)design based on neural architecture search(NAS)and knowledge distillation(KD)is proposed.It can realize the automatic construction of the space target inverse synthetic aperture radar(ISAR)image recognition model with ultra-lightweight and high accuracy.This method introduces the NAS method into the radar image recognition for the first time,which solves the time-consuming and labor-consuming problems in the artificial design of the space target ISAR image automatic recognition model(STIIARM).On this basis,the NAS model’s knowledge is transferred to the student model with lower computational complexity by the flow of the solution procedure(FSP)distillation method.Thus,the decline of recognition accuracy caused by the direct compression of model structural parameters can be effectively avoided,and the ultralightweight STIIARM can be obtained.In the method,the Inverted Linear Bottleneck(ILB)and Inverted Residual Block(IRB)are firstly taken as each block’s basic structure in CNN.And the expansion ratio,output filter size,number of IRBs,and convolution kernel size are set as the search parameters to construct a hierarchical decomposition search space.Then,the recognition accuracy and computational complexity are taken as the objective function and constraint conditions,respectively,and the global optimization model of the CNN architecture search is established.Next,the simulated annealing(SA)algorithm is used as the search strategy to search out the lightweight and high accuracy STIIARM directly.After that,based on the three principles of similar block structure,the same corresponding channel number,and the minimum computational complexity,the more lightweight student model is designed,and the FSP matrix pairing between the NAS model and student model is completed.Finally,by minimizing the loss between the FSP matrix pairs of the NAS model and student model,the student model’s weight adjustment is completed.Thus the ultra-lightweight and high accuracy STIIARM is obtained.The proposed method’s effectiveness is verified by the simulation experiments on the ISAR image dataset of five types of space targets.
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.
基金Supported by National Natural Science Foundation of China(Grant Nos.62261160575,61991414,61973036)Technical Field Foundation of the National Defense Science and Technology 173 Program of China(Grant Nos.20220601053,20220601030)。
文摘There is no unified planning standard for unstructured roads,and the morphological structures of these roads are complex and varied.It is important to maintain a balance between accuracy and speed for unstructured road extraction models.Unstructured road extraction algorithms based on deep learning have problems such as high model complexity,high computational cost,and the inability to adapt to current edge computing devices.Therefore,it is best to use lightweight network models.Considering the need for lightweight models and the characteristics of unstructured roads with different pattern shapes,such as blocks and strips,a TMB(Triple Multi-Block)feature extraction module is proposed,and the overall structure of the TMBNet network is described.The TMB module was compared with SS-nbt,Non-bottleneck-1D,and other modules via experiments.The feasibility and effectiveness of the TMB module design were proven through experiments and visualizations.The comparison experiment,using multiple convolution kernel categories,proved that the TMB module can improve the segmentation accuracy of the network.The comparison with different semantic segmentation networks demonstrates that the TMBNet network has advantages in terms of unstructured road extraction.
基金the National Natural Science Foundation of China (No.51275223)。
文摘To avoid colliding with trees during its operation,a lawn mower robot must detect the trees.Existing tree detection methods suffer from low detection accuracy(missed detection)and the lack of a lightweight model.In this study,a dataset of trees was constructed on the basis of a real lawn environment.According to the theory of channel incremental depthwise convolution and residual suppression,the Embedded-A module is proposed,which expands the depth of the feature map twice to form a residual structure to improve the lightweight degree of the model.According to residual fusion theory,the Embedded-B module is proposed,which improves the accuracy of feature-map downsampling by depthwise convolution and pooling fusion.The Embedded YOLO object detection network is formed by stacking the embedded modules and the fusion of feature maps of different resolutions.Experimental results on the testing set show that the Embedded YOLO tree detection algorithm has 84.17%and 69.91%average precision values respectively for trunk and spherical tree,and 77.04% mean average precision value.The number of convolution parameters is 1.78×10^(6),and the calculation amount is 3.85 billion float operations per second.The size of weight file is 7.11MB,and the detection speed can reach 179 frame/s.This study provides a theoretical basis for the lightweight application of the object detection algorithm based on deep learning for lawn mower robots.
