Aiming at the problems of low detection efficiency and difficult positioning of traditional steel surface defect detection methods,a lightweight steel surface defect detection model based on you only look once version...Aiming at the problems of low detection efficiency and difficult positioning of traditional steel surface defect detection methods,a lightweight steel surface defect detection model based on you only look once version 7(YOLOv7)is proposed.First,a cascading style sheets(CSS)block module is proposed,which uses more lightweight operations to obtain redundant information in the feature map,reduces the amount of computation,and effectively improves the detection speed.Secondly,the improved spatial pyramid pooling with cross stage partial convolutions(SPPCSPC)structure is adopted to ensure that the model can also pay attention to the defect location information while predicting the defect category information,obtain richer defect features.In addition,the convolution operation in the original model is simplified,which significantly reduces the size of the model and helps to improve the detection speed.Finally,using efficient intersection over union(EIOU)loss to focus on high-quality anchors,speed up convergence and improve positioning accuracy.Experiments were carried out on the Northeastern University-defect(NEU-DET)steel surface defect dataset.Compared with the original YOLOv7 model,the number of parameters of this model was reduced by 40%,the frames per second(FPS)reached 112,and the average accuracy reached 79.1%,the detection accuracy and speed have been improved,which can meet the needs of steel surface defect detection.展开更多
This study investigates the application of Learnable Memory Vision Transformers(LMViT)for detecting metal surface flaws,comparing their performance with traditional CNNs,specifically ResNet18 and ResNet50,as well as o...This study investigates the application of Learnable Memory Vision Transformers(LMViT)for detecting metal surface flaws,comparing their performance with traditional CNNs,specifically ResNet18 and ResNet50,as well as other transformer-based models including Token to Token ViT,ViT withoutmemory,and Parallel ViT.Leveraging awidely-used steel surface defect dataset,the research applies data augmentation and t-distributed stochastic neighbor embedding(t-SNE)to enhance feature extraction and understanding.These techniques mitigated overfitting,stabilized training,and improved generalization capabilities.The LMViT model achieved a test accuracy of 97.22%,significantly outperforming ResNet18(88.89%)and ResNet50(88.90%),aswell as the Token to TokenViT(88.46%),ViT without memory(87.18),and Parallel ViT(91.03%).Furthermore,LMViT exhibited superior training and validation performance,attaining a validation accuracy of 98.2%compared to 91.0%for ResNet 18,96.0%for ResNet50,and 89.12%,87.51%,and 91.21%for Token to Token ViT,ViT without memory,and Parallel ViT,respectively.The findings highlight the LMViT’s ability to capture long-range dependencies in images,an areawhere CNNs struggle due to their reliance on local receptive fields and hierarchical feature extraction.The additional transformer-based models also demonstrate improved performance in capturing complex features over CNNs,with LMViT excelling particularly at detecting subtle and complex defects,which is critical for maintaining product quality and operational efficiency in industrial applications.For instance,the LMViT model successfully identified fine scratches and minor surface irregularities that CNNs often misclassify.This study not only demonstrates LMViT’s potential for real-world defect detection but also underscores the promise of other transformer-based architectures like Token to Token ViT,ViT without memory,and Parallel ViT in industrial scenarios where complex spatial relationships are key.Future research may focus on enhancing LMViT’s computational efficiency for deployment in real-time quality control systems.展开更多
The detection of surface defects in concrete bridges using deep learning is of significant importance for reducing operational risks,saving maintenance costs,and driving the intelligent transformation of bridge defect...The detection of surface defects in concrete bridges using deep learning is of significant importance for reducing operational risks,saving maintenance costs,and driving the intelligent transformation of bridge defect detection.In contrast to the subjective and inefficient manual visual inspection,deep learning-based algorithms for concrete defect detection exhibit remarkable advantages,emerging as a focal point in recent research.This paper comprehensively analyzes the research progress of deep learning algorithms in the field of surface defect detection in concrete bridges in recent years.It introduces the early detection methods for surface defects in concrete bridges and the development of deep learning.Subsequently,it provides an overview of deep learning-based concrete bridge surface defect detection research from three aspects:image classification,object detection,and semantic segmentation.The paper summarizes the strengths and weaknesses of existing methods and the challenges they face.Additionally,it analyzes and prospects the development trends of surface defect detection in concrete bridges.展开更多
The service life of internal combustion engines is significantly influenced by surface defects in cylinder liners.To address the limitations of traditional detection methods,we propose an enhanced YOLOv8 model with Sw...The service life of internal combustion engines is significantly influenced by surface defects in cylinder liners.To address the limitations of traditional detection methods,we propose an enhanced YOLOv8 model with Swin Transformer as the backbone network.This approach leverages Swin Transformer's multi-head self-attention mechanism for improved feature extraction of defects spanning various scales.Integrated with the YOLOv8 detection head,our model achieves a mean average precision of 85.1%on our dataset,outperforming baseline methods by 1.4%.The model's effectiveness is further demonstrated on a steel-surface defect dataset,indicating its broad applicability in industrial surface defect detection.Our work highlights the potential of combining Swin Transformer and YOLOv8 for accurate and efficient defect detection.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
Defect detection based on computer vision is a critical component in ensuring the quality of industrial products.However,existing detection methods encounter several challenges in practical applications,including the ...Defect detection based on computer vision is a critical component in ensuring the quality of industrial products.However,existing detection methods encounter several challenges in practical applications,including the scarcity of labeled samples,limited adaptability of pre-trained models,and the data heterogeneity in distributed environments.To address these issues,this research proposes an unsupervised defect detection method,FLAME(Federated Learning with Adaptive Multi-Model Embeddings).The method comprises three stages:(1)Feature learning stage:this work proposes FADE(Feature-Adaptive Domain-Specific Embeddings),a framework employs Gaussian noise injection to simulate defective patterns and implements a feature discriminator for defect detection,thereby enhancing the pre-trained model’s industrial imagery representation capabilities.(2)Knowledge distillation co-training stage:a multi-model feature knowledge distillation mechanism is introduced.Through feature-level knowledge transfer between the global model and historical local models,the current local model is guided to learn better feature representations from the global model.The approach prevents local models from converging to local optima and mitigates performance degradation caused by data heterogeneity.(3)Model parameter aggregation stage:participating clients utilize weighted averaging aggregation to synthesize an updated global model,facilitating efficient knowledge consolidation.Experimental results demonstrate that FADE improves the average image-level Area under the Receiver Operating Characteristic Curve(AUROC)by 7.34%compared to methods directly utilizing pre-trained models.In federated learning environments,FLAME’s multi-model feature knowledge distillation mechanism outperforms the classic FedAvg algorithm by 2.34%in average image-level AUROC,while exhibiting superior convergence properties.展开更多
Detecting surface defects on unused rails is crucial for evaluating rail quality and durability to ensure the safety of rail transportation.However,existing detection methods often struggle with challenges such as com...Detecting surface defects on unused rails is crucial for evaluating rail quality and durability to ensure the safety of rail transportation.However,existing detection methods often struggle with challenges such as complex defect morphology,texture similarity,and fuzzy edges,leading to poor accuracy and missed detections.In order to resolve these problems,we propose MSCM-Net(Multi-Scale Cross-Modal Network),a multiscale cross-modal framework focused on detecting rail surface defects.MSCM-Net introduces an attention mechanism to dynamically weight the fusion of RGB and depth maps,effectively capturing and enhancing features at different scales for each modality.To further enrich feature representation and improve edge detection in blurred areas,we propose a multi-scale void fusion module that integrates multi-scale feature information.To improve cross-modal feature fusion,we develop a cross-enhanced fusion module that transfers fused features between layers to incorporate interlayer information.We also introduce a multimodal feature integration module,which merges modality-specific features from separate decoders into a shared decoder,enhancing detection by leveraging richer complementary information.Finally,we validate MSCM-Net on the NEU RSDDS-AUG RGB-depth dataset,comparing it against 12 leading methods,and the results show that MSCM-Net achieves superior performance on all metrics.展开更多
Roads inevitably have defects during use,which not only seriously affect their service life but also pose a hidden danger to traffic safety.Existing algorithms for detecting road defects are unsatisfactory in terms of...Roads inevitably have defects during use,which not only seriously affect their service life but also pose a hidden danger to traffic safety.Existing algorithms for detecting road defects are unsatisfactory in terms of accuracy and generalization,so this paper proposes an algorithm based on YOLOv11.The method embeds wavelet transform convolution(WTConv)into the backbone’s C3k2 module to enhance low-frequency feature extraction while avoiding parameter bloat.Secondly,a novel multi-scale fusion diffusion network(MFDN)architecture is designed for the neck to strengthen cross-scale feature interactions,boosting detection precision.In terms of model optimization,the traditional downsampling method is discarded,and the innovative Adown(adaptive downsampling)technique is adopted,which streamlines the parameter scales while effectively mitigating the information loss problem during downsampling.Finally,in this paper,we propose Wise-PIDIoU by combining WiseIoU and MPDIoU to minimize the negative impact of low-quality anchor frames and enhance the detection capability of the model.The experimental results indicate that the proposed algorithm achieves an average detection accuracy of 86.5%for mAP@50 on the RDD2022 dataset,which is 2%higher than the original algorithm while ensuring that the amount of computation is basically unchanged.The number of parameters is reduced by 17%,and the F1 score is improved by 3%,showing better detection performance than other algorithms when facing different types of defects.The excellent performance on embedded devices proves that the algorithm also has favorable application prospects in practical inspection.展开更多
Real-time detection of surface defects on cables is crucial for ensuring the safe operation of power systems.However,existing methods struggle with small target sizes,complex backgrounds,low-quality image acquisition,...Real-time detection of surface defects on cables is crucial for ensuring the safe operation of power systems.However,existing methods struggle with small target sizes,complex backgrounds,low-quality image acquisition,and interference from contamination.To address these challenges,this paper proposes the Real-time Cable Defect Detection Network(RC2DNet),which achieves an optimal balance between detection accuracy and computational efficiency.Unlike conventional approaches,RC2DNet introduces a small object feature extraction module that enhances the semantic representation of small targets through feature pyramids,multi-level feature fusion,and an adaptive weighting mechanism.Additionally,a boundary feature enhancement module is designed,incorporating boundary-aware convolution,a novel boundary attention mechanism,and an improved loss function to significantly enhance boundary localization accuracy.Experimental results demonstrate that RC2DNet outperforms state-of-the-art methods in precision,recall,F1-score,mean Intersection over Union(mIoU),and frame rate,enabling real-time and highly accurate cable defect detection in complex backgrounds.展开更多
With the rapid development of computer vision technology,artificial intelligence algorithms,and high-performance computing platforms,machine vision technology has gradually shown its great potential in automated produ...With the rapid development of computer vision technology,artificial intelligence algorithms,and high-performance computing platforms,machine vision technology has gradually shown its great potential in automated production lines,especially in defect detection.Machine vision technology can be applied in many industries such as semiconductor,automobile manufacturing,aerospace,food,and drugs,which can significantly improve detection efficiency and accuracy,reduce labor costs,improve product quality,enhance market competitiveness,and provide strong support for the arrival of Industry 4.0 era.In this article,the concept,advantages,and disadvantages of machine vision and the algorithm framework of machine vision in the defect detection system are briefly described,aiming to promote the rapid development of industry and strengthen China’s industry.展开更多
Manufacturers must identify and classify various defects in automotive sealing rings to ensure product quality.Deep learning algorithms show promise in this field,but challenges remain,especially in detecting small-sc...Manufacturers must identify and classify various defects in automotive sealing rings to ensure product quality.Deep learning algorithms show promise in this field,but challenges remain,especially in detecting small-scale defects under harsh industrial conditions with multimodal data.This paper proposes an enhanced version of You Only Look Once(YOLO)v8 for improved defect detection in automotive sealing rings.We introduce the Multi-scale Adaptive Feature Extraction(MAFE)module,which integrates Deformable ConvolutionalNetwork(DCN)and Spaceto-Depth(SPD)operations.This module effectively captures long-range dependencies,enhances spatial aggregation,and minimizes information loss of small objects during feature extraction.Furthermore,we introduce the Blur-Aware Wasserstein Distance(BAWD)loss function,which improves regression accuracy and detection capabilities for small object anchor boxes,particularly in scenarios involving defocus blur.Additionally,we have constructed a high-quality dataset of automotive sealing ring defects,providing a valuable resource for evaluating defect detection methods.Experimental results demonstrate our method’s high performance,achieving 98.30% precision,96.62% recall,and an inference speed of 20.3 ms.展开更多
High-performance lattice structures produced through powder bed fusion-laser beam exhibit high specific strength and energy absorption capabilities.However,a significant deviation exists between the mechanical propert...High-performance lattice structures produced through powder bed fusion-laser beam exhibit high specific strength and energy absorption capabilities.However,a significant deviation exists between the mechanical properties,service life of lattice structures,and design expectations.This deviation arises from the intense interaction between the laser and powder,which leads to the formation of numerous defects within the lattice structure.To address these issues,this paper proposes a high-performance defect detection model for metal lattice structures based on YOLOv4,called YOLO-Lattice(YOLO-L).The main objectives of this paper are as follows:(1)utilize computed tomography to construct datasets of the diamond lattice and body-centered cubic lattice structures;(2)in the backbone network of YOLOv4,employ deformable convolution to enhance the feature extraction capability of the model for small-scale defects;(3)adopt a dual-attention mechanism to suppress invalid feature information and amplify the distinction between defect and background regions;and(4)implement a channel pruning strategy to eliminate channels carrying less feature information,thereby improving the inference speed of the model.The experimental results on the diamond lattice structure dataset demonstrate that the mean average precision of the YOLO-L model increased from 96.98% to 98.8%(with an intersection over union of 0.5),and the inference speed decreased from 51.3 ms to 32.5 ms when compared to YOLOv4.Thus,the YOLO-L model can be effectively used to detect defects in metal lattice structures.展开更多
To solve the problems of the low accuracy and poor real-time performance of traditional strip steel surface defect detection meth-ods,which are caused by the characteristics of many kinds,complex shapes,and different ...To solve the problems of the low accuracy and poor real-time performance of traditional strip steel surface defect detection meth-ods,which are caused by the characteristics of many kinds,complex shapes,and different scales of strip surface defects,a strip steel surface defect detection algorithm based on improved Faster R-CNN is proposed.Firstly,the residual convolution module is inserted into the Swin Transformer network module to form the RC-Swin Transformer network module,and the RC-Swin Transformer module is introduced into the backbone network of the traditional Faster R-CNN to enhance the ability of the network to extract the global feature information of the image and adapt to the complex shape of the strip steel surface defect.To improve the attention of the network to defects in the image,a CBAM-BiFPN network module is designed,and then the backbone network is combined with the CBAM-BiFPN network to realize the de-tection and fusion of multi-scale features.The RoI align layer is used instead of the RoI pooling layer to improve the accuracy of defect loca-tion.Finally,Soft NMS is used to achieve non-maximum suppression and remove redundant boxes.In the comparative experiment on the NEU-DET dataset,the improved algorithm improves the mean average precision by 4.2%compared with the Faster R-CNN algorithm,and also improves the average precision by 6.1%and 6.7%for crazing defect and rolled-in scale defect,which are difficult to detect with the Faster R-CNN algorithm.The experiments show that the improvements proposed in the paper effectively improve the detection accuracy of the algorithm and have certain practical value.展开更多
Urban underground pipelines are an important infrastructure in cities,and timely investigation of problems in underground pipelines can help ensure the normal operation of cities.Owing to the growing demand for defect...Urban underground pipelines are an important infrastructure in cities,and timely investigation of problems in underground pipelines can help ensure the normal operation of cities.Owing to the growing demand for defect detection in urban underground pipelines,this study developed an improved defect detection method for urban underground pipelines based on fully convolutional one-stage object detector(FCOS),called spatial pyramid pooling-fast(SPPF)feature fusion and dual detection heads based on FCOS(SDH-FCOS)model.This study improved the feature fusion component of the model network based on FCOS,introduced an SPPF network structure behind the last output feature layer of the backbone network,fused the local and global features,added a top-down path to accelerate the circulation of shallowinformation,and enriched the semantic information acquired by shallow features.The ability of the model to detect objects with multiple morphologies was strengthened by introducing dual detection heads.The experimental results using an open dataset of underground pipes show that the proposed SDH-FCOS model can recognize underground pipe defects more accurately;the average accuracy was improved by 2.7% compared with the original FCOS model,reducing the leakage rate to a large extent and achieving real-time detection.Also,our model achieved a good trade-off between accuracy and speed compared with other mainstream methods.This proved the effectiveness of the proposed model.展开更多
As computer technology continues to advance,factories have increasingly higher demands for detecting defects.However,detecting defects in a plant environment remains a challenging task due to the presence of complex b...As computer technology continues to advance,factories have increasingly higher demands for detecting defects.However,detecting defects in a plant environment remains a challenging task due to the presence of complex backgrounds and defects of varying shapes and sizes.To address this issue,this paper proposes YOLO-DD,a defect detectionmodel based on YOLOv5 that is effective and robust.To improve the feature extraction process and better capture global information,the vanilla YOLOv5 is augmented with a new module called Relative-Distance-Aware Transformer(RDAT).Additionally,an Information Gap Filling Strategy(IGFS)is proposed to improve the fusion of features at different scales.The classic lightweight attention mechanism Squeeze-and-Excitation(SE)module is also incorporated into the neck section to enhance feature expression and improve the model’s performance.Experimental results on the NEU-DET dataset demonstrate that YOLO-DDachieves competitive results compared to state-of-the-art methods,with a 2.0% increase in accuracy compared to the original YOLOv5,achieving 82.41% accuracy and38.25FPS(framesper second).Themodel is also testedon a self-constructed fabric defect dataset,and the results show that YOLO-DD is more stable and has higher accuracy than the original YOLOv5,demonstrating its stability and generalization ability.The high efficiency of YOLO-DD enables it to meet the requirements of industrial high accuracy and real-time detection.展开更多
Timely inspection of defects on the surfaces of wind turbine blades can effectively prevent unpredictable accidents.To this end,this study proposes a semi-supervised object-detection network based on You Only Looking ...Timely inspection of defects on the surfaces of wind turbine blades can effectively prevent unpredictable accidents.To this end,this study proposes a semi-supervised object-detection network based on You Only Looking Once version 4(YOLOv4).A semi-supervised structure comprising a generative adversarial network(GAN)was designed to overcome the difficulty in obtaining sufficient samples and sample labeling.In a GAN,the generator is realized by an encoder-decoder network,where the backbone of the encoder is YOLOv4 and the decoder comprises inverse convolutional layers.Partial features from the generator are passed to the defect detection network.Deploying several unlabeled images can significantly improve the generalization and recognition capabilities of defect-detection models.The small-scale object detection capacity of the network can be improved by enhancing essential features in the feature map by adding the concurrent spatial and channel squeeze and excitation(scSE)attention module to the three parts of the YOLOv4 network.A balancing improvement was made to the loss function of YOLOv4 to overcome the imbalance problem of the defective species.The results for both the single-and multi-category defect datasets show that the improved model can make good use of the features of the unlabeled images.The accuracy of wind turbine blade defect detection also has a significant advantage over classical object detection algorithms,including faster R-CNN and DETR.展开更多
Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal depende...Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal dependence,and noise.Therefore,methodologies for data augmentation and conversion of time series data into images for analysis have been studied.This paper proposes a fault detection model that uses time series data augmentation and transformation to address the problems of data imbalance,temporal dependence,and robustness to noise.The method of data augmentation is set as the addition of noise.It involves adding Gaussian noise,with the noise level set to 0.002,to maximize the generalization performance of the model.In addition,we use the Markov Transition Field(MTF)method to effectively visualize the dynamic transitions of the data while converting the time series data into images.It enables the identification of patterns in time series data and assists in capturing the sequential dependencies of the data.For anomaly detection,the PatchCore model is applied to show excellent performance,and the detected anomaly areas are represented as heat maps.It allows for the detection of anomalies,and by applying an anomaly map to the original image,it is possible to capture the areas where anomalies occur.The performance evaluation shows that both F1-score and Accuracy are high when time series data is converted to images.Additionally,when processed as images rather than as time series data,there was a significant reduction in both the size of the data and the training time.The proposed method can provide an important springboard for research in the field of anomaly detection using time series data.Besides,it helps solve problems such as analyzing complex patterns in data lightweight.展开更多
Printed Circuit Board(PCB)surface tiny defect detection is a difficult task in the integrated circuit industry,especially since the detection of tiny defects on PCB boards with large-size complex circuits has become o...Printed Circuit Board(PCB)surface tiny defect detection is a difficult task in the integrated circuit industry,especially since the detection of tiny defects on PCB boards with large-size complex circuits has become one of the bottlenecks.To improve the performance of PCB surface tiny defects detection,a PCB tiny defects detection model based on an improved attention residual network(YOLOX-AttResNet)is proposed.First,the unsupervised clustering performance of the K-means algorithm is exploited to optimize the channel weights for subsequent operations by feeding the feature mapping into the SENet(Squeeze and Excitation Network)attention network;then the improved K-means-SENet network is fused with the directly mapped edges of the traditional ResNet network to form an augmented residual network(AttResNet);and finally,the AttResNet module is substituted for the traditional ResNet structure in the backbone feature extraction network of mainstream excellent detection models,thus improving the ability to extract small features from the backbone of the target detection network.The results of ablation experiments on a PCB surface defect dataset show that AttResNet is a reliable and efficient module.In Torify the performance of AttResNet for detecting small defects in large-size complex circuit images,a series of comparison experiments are further performed.The results show that the AttResNet module combines well with the five best existing target detection frameworks(YOLOv3,YOLOX,Faster R-CNN,TDD-Net,Cascade R-CNN),and all the combined new models have improved detection accuracy compared to the original model,which suggests that the AttResNet module proposed in this paper can help the detection model to extract target features.Among them,the YOLOX-AttResNet model proposed in this paper performs the best,with the highest accuracy of 98.45% and the detection speed of 36 FPS(Frames Per Second),which meets the accuracy and real-time requirements for the detection of tiny defects on PCB surfaces.This study can provide some new ideas for other real-time online detection tasks of tiny targets with high-resolution images.展开更多
To advance the printing manufacturing industry towards intelligence and address the challenges faced by supervised learning,such as the high workload,cost,poor generalization,and labeling issues,an unsupervised and tr...To advance the printing manufacturing industry towards intelligence and address the challenges faced by supervised learning,such as the high workload,cost,poor generalization,and labeling issues,an unsupervised and transfer learning-based method for printing defect detection was proposed in this study.This method enabled defect detection in printed surface without the need for extensive labeled defect.The ResNet101-SSTU model was used in this study.On the public dataset of printing defect images,the ResNet101-SSTU model not only achieves comparable performance and speed to mainstream supervised learning detection models but also successfully addresses some of the detection challenges encountered in supervised learning.The proposed ResNet101-SSTU model effectively eliminates the need for extensive defect samples and labeled data in training,providing an efficient solution for quality inspection in the printing industry.展开更多
Defect detection technology is crucial for the efficient operation and maintenance of photovoltaic systems.However,the diversity of defect types,scale inconsistencies,and background interference significantly complica...Defect detection technology is crucial for the efficient operation and maintenance of photovoltaic systems.However,the diversity of defect types,scale inconsistencies,and background interference significantly complicate the detection task.Therefore,this paper employs the YOLOX model as the backbone network structure and optimizes various modules to address these issues.First,we adopt a transfer learning strategy to accelerate model convergence and avoid insufficient accuracy due to the limited number of defect samples.Second,we introduce the SENet module into the feature extraction process to enhance the contrast between defects and their background.Then,we incorporate the ASFF strategy at the end of the PAFPN network to adaptively learn and emphasize both high-and low-level semantic features of defects.Furthermore,model accuracy is enhanced by refining the loss functions for positioning,classification,and confidence.Finally,the proposed method achieved excellent results on the Photovoltaic Electroluminescence Anomaly Detection dataset(PVEL-AD),with a mAP of 96.7%and a detection speed of 71.47FPS.Specifically,the detection of small target defects showed significant improvement.展开更多
基金supported by the National Natural Science Foundation of China(No.62103298)the Natural Science Foundation of Hebei Province(No.F2018209289)。
文摘Aiming at the problems of low detection efficiency and difficult positioning of traditional steel surface defect detection methods,a lightweight steel surface defect detection model based on you only look once version 7(YOLOv7)is proposed.First,a cascading style sheets(CSS)block module is proposed,which uses more lightweight operations to obtain redundant information in the feature map,reduces the amount of computation,and effectively improves the detection speed.Secondly,the improved spatial pyramid pooling with cross stage partial convolutions(SPPCSPC)structure is adopted to ensure that the model can also pay attention to the defect location information while predicting the defect category information,obtain richer defect features.In addition,the convolution operation in the original model is simplified,which significantly reduces the size of the model and helps to improve the detection speed.Finally,using efficient intersection over union(EIOU)loss to focus on high-quality anchors,speed up convergence and improve positioning accuracy.Experiments were carried out on the Northeastern University-defect(NEU-DET)steel surface defect dataset.Compared with the original YOLOv7 model,the number of parameters of this model was reduced by 40%,the frames per second(FPS)reached 112,and the average accuracy reached 79.1%,the detection accuracy and speed have been improved,which can meet the needs of steel surface defect detection.
基金funded by Woosong University Academic Research 2024.
文摘This study investigates the application of Learnable Memory Vision Transformers(LMViT)for detecting metal surface flaws,comparing their performance with traditional CNNs,specifically ResNet18 and ResNet50,as well as other transformer-based models including Token to Token ViT,ViT withoutmemory,and Parallel ViT.Leveraging awidely-used steel surface defect dataset,the research applies data augmentation and t-distributed stochastic neighbor embedding(t-SNE)to enhance feature extraction and understanding.These techniques mitigated overfitting,stabilized training,and improved generalization capabilities.The LMViT model achieved a test accuracy of 97.22%,significantly outperforming ResNet18(88.89%)and ResNet50(88.90%),aswell as the Token to TokenViT(88.46%),ViT without memory(87.18),and Parallel ViT(91.03%).Furthermore,LMViT exhibited superior training and validation performance,attaining a validation accuracy of 98.2%compared to 91.0%for ResNet 18,96.0%for ResNet50,and 89.12%,87.51%,and 91.21%for Token to Token ViT,ViT without memory,and Parallel ViT,respectively.The findings highlight the LMViT’s ability to capture long-range dependencies in images,an areawhere CNNs struggle due to their reliance on local receptive fields and hierarchical feature extraction.The additional transformer-based models also demonstrate improved performance in capturing complex features over CNNs,with LMViT excelling particularly at detecting subtle and complex defects,which is critical for maintaining product quality and operational efficiency in industrial applications.For instance,the LMViT model successfully identified fine scratches and minor surface irregularities that CNNs often misclassify.This study not only demonstrates LMViT’s potential for real-world defect detection but also underscores the promise of other transformer-based architectures like Token to Token ViT,ViT without memory,and Parallel ViT in industrial scenarios where complex spatial relationships are key.Future research may focus on enhancing LMViT’s computational efficiency for deployment in real-time quality control systems.
基金supported by the Key Research and Development Program of Shaanxi Province-International Science and Technology Cooperation Program Project (No.2020KW-001)the Contract for Xi'an Municipal Science and Technology Plan Project-Xi'an City Strong Foundation Innovation Plan (No.21XJZZ0074)the Key Project of Graduate Student Innovation Fund at Xi'an University of Posts and Telecommunications (No.CXJJZL2023013)。
文摘The detection of surface defects in concrete bridges using deep learning is of significant importance for reducing operational risks,saving maintenance costs,and driving the intelligent transformation of bridge defect detection.In contrast to the subjective and inefficient manual visual inspection,deep learning-based algorithms for concrete defect detection exhibit remarkable advantages,emerging as a focal point in recent research.This paper comprehensively analyzes the research progress of deep learning algorithms in the field of surface defect detection in concrete bridges in recent years.It introduces the early detection methods for surface defects in concrete bridges and the development of deep learning.Subsequently,it provides an overview of deep learning-based concrete bridge surface defect detection research from three aspects:image classification,object detection,and semantic segmentation.The paper summarizes the strengths and weaknesses of existing methods and the challenges they face.Additionally,it analyzes and prospects the development trends of surface defect detection in concrete bridges.
基金supported by the Scientific and technological key project in Henan Province 22210224002the Natural Science Foundation of Henan Polytechnic University B2021-38.
文摘The service life of internal combustion engines is significantly influenced by surface defects in cylinder liners.To address the limitations of traditional detection methods,we propose an enhanced YOLOv8 model with Swin Transformer as the backbone network.This approach leverages Swin Transformer's multi-head self-attention mechanism for improved feature extraction of defects spanning various scales.Integrated with the YOLOv8 detection head,our model achieves a mean average precision of 85.1%on our dataset,outperforming baseline methods by 1.4%.The model's effectiveness is further demonstrated on a steel-surface defect dataset,indicating its broad applicability in industrial surface defect detection.Our work highlights the potential of combining Swin Transformer and YOLOv8 for accurate and efficient defect detection.
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 in part by the National Natural Science Foundation of China under Grants 32171909,52205254,32301704the Guangdong Basic and Applied Basic Research Foundation under Grants 2023A1515011255,2024A1515010199+1 种基金the Scientific Research Projects of Universities in Guangdong Province under Grants 2024ZDZX1042,2024ZDZX3057the Ji-Hua Laboratory Open Project under Grant X220931UZ230.
文摘Defect detection based on computer vision is a critical component in ensuring the quality of industrial products.However,existing detection methods encounter several challenges in practical applications,including the scarcity of labeled samples,limited adaptability of pre-trained models,and the data heterogeneity in distributed environments.To address these issues,this research proposes an unsupervised defect detection method,FLAME(Federated Learning with Adaptive Multi-Model Embeddings).The method comprises three stages:(1)Feature learning stage:this work proposes FADE(Feature-Adaptive Domain-Specific Embeddings),a framework employs Gaussian noise injection to simulate defective patterns and implements a feature discriminator for defect detection,thereby enhancing the pre-trained model’s industrial imagery representation capabilities.(2)Knowledge distillation co-training stage:a multi-model feature knowledge distillation mechanism is introduced.Through feature-level knowledge transfer between the global model and historical local models,the current local model is guided to learn better feature representations from the global model.The approach prevents local models from converging to local optima and mitigates performance degradation caused by data heterogeneity.(3)Model parameter aggregation stage:participating clients utilize weighted averaging aggregation to synthesize an updated global model,facilitating efficient knowledge consolidation.Experimental results demonstrate that FADE improves the average image-level Area under the Receiver Operating Characteristic Curve(AUROC)by 7.34%compared to methods directly utilizing pre-trained models.In federated learning environments,FLAME’s multi-model feature knowledge distillation mechanism outperforms the classic FedAvg algorithm by 2.34%in average image-level AUROC,while exhibiting superior convergence properties.
基金funded by the National Natural Science Foundation of China(grant number 62306186)the Technology Plan Joint Foundation of Liaoning Province(grant number 2023-MSLH-246)the Technology Plan Joint Foundation of Liaoning Province(grant number 2023-BSBA-238).
文摘Detecting surface defects on unused rails is crucial for evaluating rail quality and durability to ensure the safety of rail transportation.However,existing detection methods often struggle with challenges such as complex defect morphology,texture similarity,and fuzzy edges,leading to poor accuracy and missed detections.In order to resolve these problems,we propose MSCM-Net(Multi-Scale Cross-Modal Network),a multiscale cross-modal framework focused on detecting rail surface defects.MSCM-Net introduces an attention mechanism to dynamically weight the fusion of RGB and depth maps,effectively capturing and enhancing features at different scales for each modality.To further enrich feature representation and improve edge detection in blurred areas,we propose a multi-scale void fusion module that integrates multi-scale feature information.To improve cross-modal feature fusion,we develop a cross-enhanced fusion module that transfers fused features between layers to incorporate interlayer information.We also introduce a multimodal feature integration module,which merges modality-specific features from separate decoders into a shared decoder,enhancing detection by leveraging richer complementary information.Finally,we validate MSCM-Net on the NEU RSDDS-AUG RGB-depth dataset,comparing it against 12 leading methods,and the results show that MSCM-Net achieves superior performance on all metrics.
文摘Roads inevitably have defects during use,which not only seriously affect their service life but also pose a hidden danger to traffic safety.Existing algorithms for detecting road defects are unsatisfactory in terms of accuracy and generalization,so this paper proposes an algorithm based on YOLOv11.The method embeds wavelet transform convolution(WTConv)into the backbone’s C3k2 module to enhance low-frequency feature extraction while avoiding parameter bloat.Secondly,a novel multi-scale fusion diffusion network(MFDN)architecture is designed for the neck to strengthen cross-scale feature interactions,boosting detection precision.In terms of model optimization,the traditional downsampling method is discarded,and the innovative Adown(adaptive downsampling)technique is adopted,which streamlines the parameter scales while effectively mitigating the information loss problem during downsampling.Finally,in this paper,we propose Wise-PIDIoU by combining WiseIoU and MPDIoU to minimize the negative impact of low-quality anchor frames and enhance the detection capability of the model.The experimental results indicate that the proposed algorithm achieves an average detection accuracy of 86.5%for mAP@50 on the RDD2022 dataset,which is 2%higher than the original algorithm while ensuring that the amount of computation is basically unchanged.The number of parameters is reduced by 17%,and the F1 score is improved by 3%,showing better detection performance than other algorithms when facing different types of defects.The excellent performance on embedded devices proves that the algorithm also has favorable application prospects in practical inspection.
基金supported by the National Natural Science Foundation of China under Grant 62306128the Basic Science Research Project of Jiangsu Provincial Department of Education under Grant 23KJD520003the Leading Innovation Project of Changzhou Science and Technology Bureau under Grant CQ20230072.
文摘Real-time detection of surface defects on cables is crucial for ensuring the safe operation of power systems.However,existing methods struggle with small target sizes,complex backgrounds,low-quality image acquisition,and interference from contamination.To address these challenges,this paper proposes the Real-time Cable Defect Detection Network(RC2DNet),which achieves an optimal balance between detection accuracy and computational efficiency.Unlike conventional approaches,RC2DNet introduces a small object feature extraction module that enhances the semantic representation of small targets through feature pyramids,multi-level feature fusion,and an adaptive weighting mechanism.Additionally,a boundary feature enhancement module is designed,incorporating boundary-aware convolution,a novel boundary attention mechanism,and an improved loss function to significantly enhance boundary localization accuracy.Experimental results demonstrate that RC2DNet outperforms state-of-the-art methods in precision,recall,F1-score,mean Intersection over Union(mIoU),and frame rate,enabling real-time and highly accurate cable defect detection in complex backgrounds.
文摘With the rapid development of computer vision technology,artificial intelligence algorithms,and high-performance computing platforms,machine vision technology has gradually shown its great potential in automated production lines,especially in defect detection.Machine vision technology can be applied in many industries such as semiconductor,automobile manufacturing,aerospace,food,and drugs,which can significantly improve detection efficiency and accuracy,reduce labor costs,improve product quality,enhance market competitiveness,and provide strong support for the arrival of Industry 4.0 era.In this article,the concept,advantages,and disadvantages of machine vision and the algorithm framework of machine vision in the defect detection system are briefly described,aiming to promote the rapid development of industry and strengthen China’s industry.
文摘Manufacturers must identify and classify various defects in automotive sealing rings to ensure product quality.Deep learning algorithms show promise in this field,but challenges remain,especially in detecting small-scale defects under harsh industrial conditions with multimodal data.This paper proposes an enhanced version of You Only Look Once(YOLO)v8 for improved defect detection in automotive sealing rings.We introduce the Multi-scale Adaptive Feature Extraction(MAFE)module,which integrates Deformable ConvolutionalNetwork(DCN)and Spaceto-Depth(SPD)operations.This module effectively captures long-range dependencies,enhances spatial aggregation,and minimizes information loss of small objects during feature extraction.Furthermore,we introduce the Blur-Aware Wasserstein Distance(BAWD)loss function,which improves regression accuracy and detection capabilities for small object anchor boxes,particularly in scenarios involving defocus blur.Additionally,we have constructed a high-quality dataset of automotive sealing ring defects,providing a valuable resource for evaluating defect detection methods.Experimental results demonstrate our method’s high performance,achieving 98.30% precision,96.62% recall,and an inference speed of 20.3 ms.
基金supported by Natural Science Foundation of China(Grant No.52175488)Scientific Research Program for Young Outstanding Talent of Higher Education of Hebei Province(China)(Grant No.BJ2021045)S&T Program of Hebei(China)(Grant No.236Z1808G).
文摘High-performance lattice structures produced through powder bed fusion-laser beam exhibit high specific strength and energy absorption capabilities.However,a significant deviation exists between the mechanical properties,service life of lattice structures,and design expectations.This deviation arises from the intense interaction between the laser and powder,which leads to the formation of numerous defects within the lattice structure.To address these issues,this paper proposes a high-performance defect detection model for metal lattice structures based on YOLOv4,called YOLO-Lattice(YOLO-L).The main objectives of this paper are as follows:(1)utilize computed tomography to construct datasets of the diamond lattice and body-centered cubic lattice structures;(2)in the backbone network of YOLOv4,employ deformable convolution to enhance the feature extraction capability of the model for small-scale defects;(3)adopt a dual-attention mechanism to suppress invalid feature information and amplify the distinction between defect and background regions;and(4)implement a channel pruning strategy to eliminate channels carrying less feature information,thereby improving the inference speed of the model.The experimental results on the diamond lattice structure dataset demonstrate that the mean average precision of the YOLO-L model increased from 96.98% to 98.8%(with an intersection over union of 0.5),and the inference speed decreased from 51.3 ms to 32.5 ms when compared to YOLOv4.Thus,the YOLO-L model can be effectively used to detect defects in metal lattice structures.
基金supported by the National Natural Science Foundation of China(12002138).
文摘To solve the problems of the low accuracy and poor real-time performance of traditional strip steel surface defect detection meth-ods,which are caused by the characteristics of many kinds,complex shapes,and different scales of strip surface defects,a strip steel surface defect detection algorithm based on improved Faster R-CNN is proposed.Firstly,the residual convolution module is inserted into the Swin Transformer network module to form the RC-Swin Transformer network module,and the RC-Swin Transformer module is introduced into the backbone network of the traditional Faster R-CNN to enhance the ability of the network to extract the global feature information of the image and adapt to the complex shape of the strip steel surface defect.To improve the attention of the network to defects in the image,a CBAM-BiFPN network module is designed,and then the backbone network is combined with the CBAM-BiFPN network to realize the de-tection and fusion of multi-scale features.The RoI align layer is used instead of the RoI pooling layer to improve the accuracy of defect loca-tion.Finally,Soft NMS is used to achieve non-maximum suppression and remove redundant boxes.In the comparative experiment on the NEU-DET dataset,the improved algorithm improves the mean average precision by 4.2%compared with the Faster R-CNN algorithm,and also improves the average precision by 6.1%and 6.7%for crazing defect and rolled-in scale defect,which are difficult to detect with the Faster R-CNN algorithm.The experiments show that the improvements proposed in the paper effectively improve the detection accuracy of the algorithm and have certain practical value.
基金supported by the National Natural Science Foundation of China under Grant No.61976226the Research and Academic Team of South-CentralMinzu University under Grant No.KTZ20050.
文摘Urban underground pipelines are an important infrastructure in cities,and timely investigation of problems in underground pipelines can help ensure the normal operation of cities.Owing to the growing demand for defect detection in urban underground pipelines,this study developed an improved defect detection method for urban underground pipelines based on fully convolutional one-stage object detector(FCOS),called spatial pyramid pooling-fast(SPPF)feature fusion and dual detection heads based on FCOS(SDH-FCOS)model.This study improved the feature fusion component of the model network based on FCOS,introduced an SPPF network structure behind the last output feature layer of the backbone network,fused the local and global features,added a top-down path to accelerate the circulation of shallowinformation,and enriched the semantic information acquired by shallow features.The ability of the model to detect objects with multiple morphologies was strengthened by introducing dual detection heads.The experimental results using an open dataset of underground pipes show that the proposed SDH-FCOS model can recognize underground pipe defects more accurately;the average accuracy was improved by 2.7% compared with the original FCOS model,reducing the leakage rate to a large extent and achieving real-time detection.Also,our model achieved a good trade-off between accuracy and speed compared with other mainstream methods.This proved the effectiveness of the proposed model.
基金supported in part by the National Natural Science Foundation of China under Grants 32171909,51705365,52205254The Guangdong Basic and Applied Basic Research Foundation under Grants 2020B1515120050,2023A1515011255+2 种基金The Guangdong Key R&D projects under Grant 2020B0404030001the Scientific Research Projects of Universities in Guangdong Province under Grant 2020KCXTD015The Ji Hua Laboratory Open Project under Grant X220931UZ230.
文摘As computer technology continues to advance,factories have increasingly higher demands for detecting defects.However,detecting defects in a plant environment remains a challenging task due to the presence of complex backgrounds and defects of varying shapes and sizes.To address this issue,this paper proposes YOLO-DD,a defect detectionmodel based on YOLOv5 that is effective and robust.To improve the feature extraction process and better capture global information,the vanilla YOLOv5 is augmented with a new module called Relative-Distance-Aware Transformer(RDAT).Additionally,an Information Gap Filling Strategy(IGFS)is proposed to improve the fusion of features at different scales.The classic lightweight attention mechanism Squeeze-and-Excitation(SE)module is also incorporated into the neck section to enhance feature expression and improve the model’s performance.Experimental results on the NEU-DET dataset demonstrate that YOLO-DDachieves competitive results compared to state-of-the-art methods,with a 2.0% increase in accuracy compared to the original YOLOv5,achieving 82.41% accuracy and38.25FPS(framesper second).Themodel is also testedon a self-constructed fabric defect dataset,and the results show that YOLO-DD is more stable and has higher accuracy than the original YOLOv5,demonstrating its stability and generalization ability.The high efficiency of YOLO-DD enables it to meet the requirements of industrial high accuracy and real-time detection.
基金supported in part by the National Natural Science Foundation of China under grants 62202044 and 62372039Scientific and Technological Innovation Foundation of Foshan under grant BK22BF009+3 种基金Excellent Youth Team Project for the Central Universities under grant FRF-EYIT-23-01Fundamental Research Funds for the Central Universities under grants 06500103 and 06500078Guangdong Basic and Applied Basic Research Foundation under grant 2022A1515240044Beijing Natural Science Foundation under grant 4232040.
文摘Timely inspection of defects on the surfaces of wind turbine blades can effectively prevent unpredictable accidents.To this end,this study proposes a semi-supervised object-detection network based on You Only Looking Once version 4(YOLOv4).A semi-supervised structure comprising a generative adversarial network(GAN)was designed to overcome the difficulty in obtaining sufficient samples and sample labeling.In a GAN,the generator is realized by an encoder-decoder network,where the backbone of the encoder is YOLOv4 and the decoder comprises inverse convolutional layers.Partial features from the generator are passed to the defect detection network.Deploying several unlabeled images can significantly improve the generalization and recognition capabilities of defect-detection models.The small-scale object detection capacity of the network can be improved by enhancing essential features in the feature map by adding the concurrent spatial and channel squeeze and excitation(scSE)attention module to the three parts of the YOLOv4 network.A balancing improvement was made to the loss function of YOLOv4 to overcome the imbalance problem of the defective species.The results for both the single-and multi-category defect datasets show that the improved model can make good use of the features of the unlabeled images.The accuracy of wind turbine blade defect detection also has a significant advantage over classical object detection algorithms,including faster R-CNN and DETR.
基金This research was financially supported by the Ministry of Trade,Industry,and Energy(MOTIE),Korea,under the“Project for Research and Development with Middle Markets Enterprises and DNA(Data,Network,AI)Universities”(AI-based Safety Assessment and Management System for Concrete Structures)(ReferenceNumber P0024559)supervised by theKorea Institute for Advancement of Technology(KIAT).
文摘Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal dependence,and noise.Therefore,methodologies for data augmentation and conversion of time series data into images for analysis have been studied.This paper proposes a fault detection model that uses time series data augmentation and transformation to address the problems of data imbalance,temporal dependence,and robustness to noise.The method of data augmentation is set as the addition of noise.It involves adding Gaussian noise,with the noise level set to 0.002,to maximize the generalization performance of the model.In addition,we use the Markov Transition Field(MTF)method to effectively visualize the dynamic transitions of the data while converting the time series data into images.It enables the identification of patterns in time series data and assists in capturing the sequential dependencies of the data.For anomaly detection,the PatchCore model is applied to show excellent performance,and the detected anomaly areas are represented as heat maps.It allows for the detection of anomalies,and by applying an anomaly map to the original image,it is possible to capture the areas where anomalies occur.The performance evaluation shows that both F1-score and Accuracy are high when time series data is converted to images.Additionally,when processed as images rather than as time series data,there was a significant reduction in both the size of the data and the training time.The proposed method can provide an important springboard for research in the field of anomaly detection using time series data.Besides,it helps solve problems such as analyzing complex patterns in data lightweight.
基金supported by the National Natural Science Foundation of China(No.61976083)Hubei Province Key R&D Program of China(No.2022BBA0016).
文摘Printed Circuit Board(PCB)surface tiny defect detection is a difficult task in the integrated circuit industry,especially since the detection of tiny defects on PCB boards with large-size complex circuits has become one of the bottlenecks.To improve the performance of PCB surface tiny defects detection,a PCB tiny defects detection model based on an improved attention residual network(YOLOX-AttResNet)is proposed.First,the unsupervised clustering performance of the K-means algorithm is exploited to optimize the channel weights for subsequent operations by feeding the feature mapping into the SENet(Squeeze and Excitation Network)attention network;then the improved K-means-SENet network is fused with the directly mapped edges of the traditional ResNet network to form an augmented residual network(AttResNet);and finally,the AttResNet module is substituted for the traditional ResNet structure in the backbone feature extraction network of mainstream excellent detection models,thus improving the ability to extract small features from the backbone of the target detection network.The results of ablation experiments on a PCB surface defect dataset show that AttResNet is a reliable and efficient module.In Torify the performance of AttResNet for detecting small defects in large-size complex circuit images,a series of comparison experiments are further performed.The results show that the AttResNet module combines well with the five best existing target detection frameworks(YOLOv3,YOLOX,Faster R-CNN,TDD-Net,Cascade R-CNN),and all the combined new models have improved detection accuracy compared to the original model,which suggests that the AttResNet module proposed in this paper can help the detection model to extract target features.Among them,the YOLOX-AttResNet model proposed in this paper performs the best,with the highest accuracy of 98.45% and the detection speed of 36 FPS(Frames Per Second),which meets the accuracy and real-time requirements for the detection of tiny defects on PCB surfaces.This study can provide some new ideas for other real-time online detection tasks of tiny targets with high-resolution images.
文摘To advance the printing manufacturing industry towards intelligence and address the challenges faced by supervised learning,such as the high workload,cost,poor generalization,and labeling issues,an unsupervised and transfer learning-based method for printing defect detection was proposed in this study.This method enabled defect detection in printed surface without the need for extensive labeled defect.The ResNet101-SSTU model was used in this study.On the public dataset of printing defect images,the ResNet101-SSTU model not only achieves comparable performance and speed to mainstream supervised learning detection models but also successfully addresses some of the detection challenges encountered in supervised learning.The proposed ResNet101-SSTU model effectively eliminates the need for extensive defect samples and labeled data in training,providing an efficient solution for quality inspection in the printing industry.
基金supported by the National Natural Science Foundation of China under Grant 62266034the Ningxia Natural Science Foundation Key Program underGrant2023AAC02011.
文摘Defect detection technology is crucial for the efficient operation and maintenance of photovoltaic systems.However,the diversity of defect types,scale inconsistencies,and background interference significantly complicate the detection task.Therefore,this paper employs the YOLOX model as the backbone network structure and optimizes various modules to address these issues.First,we adopt a transfer learning strategy to accelerate model convergence and avoid insufficient accuracy due to the limited number of defect samples.Second,we introduce the SENet module into the feature extraction process to enhance the contrast between defects and their background.Then,we incorporate the ASFF strategy at the end of the PAFPN network to adaptively learn and emphasize both high-and low-level semantic features of defects.Furthermore,model accuracy is enhanced by refining the loss functions for positioning,classification,and confidence.Finally,the proposed method achieved excellent results on the Photovoltaic Electroluminescence Anomaly Detection dataset(PVEL-AD),with a mAP of 96.7%and a detection speed of 71.47FPS.Specifically,the detection of small target defects showed significant improvement.
