Defect Detection Of Network Transformer Based On Improved YOVO V3

At present,most of the research on network transformer visual inspection is print and scratch detection based on machine vision,but the detection of small defects in network transformer is relatively less.Therefore,it still depends on inefficient and labor-cost detection.Based on the above reasons,the internal defect detection of network transformer is taken as the research object.Aiming at the detection difficulties caused by the small proportion of internal defects and the difficulty in obtaining effective features,improved the YOLO v3target detection algorithm to realize the efficient and accurate detection of internal defects of network transformers.Main research work is as follows:(1)The defect sample set is formed by collecting limited defect samples of network transformers and manually manufacturing defects to normal network transformers.At the same time,the defect image set is extended by rotating,mirroring and increasing Gaussian noise.Finally,the dataset is annotated by label Img to form the network transformer defect detection dataset.(2)In view of the the defects of network transformer are similar to some background textures which is difficult to extract effective features,proposed a new channel and spatial attention module to improve the ability of network to extract defect features.Based on the theoretical basis of SE attention module,the proposed channel attention module(PAM)reduces the new parameters from?(c~2)to?(c)and maintains a better direct correlation mapping relationship compared with the SE channel attention module.At the same time,added the spatial attention module(SAM),which reduces the influence of the background in the space and enhances ability to extract spatial features with a small number of parameters.(3)According to the characteristics of small proportion of defect size,improved the YOLO v3 network structure.In the input layer,used a higher resolution(608×608)to improve the detection accuracy of small targets.In the backbone network,by using Ghost lightweight convolution module to replace the ordinary convolution to obtain more computing space and more network optimization,meanwhile the attention module is used to embed the backbone network to make up for the accuracy.In connection layer,using improved PAN structure to better deep and shallow layer information fusion.In the detection layer,add the detection of 4×downsampling fusion feature as the main detection head of small target detection,meanwhile the detection of 8×down-sampling layer and 16×down-sampling feature is removed to reduce calculation of high resolution feature map detection.At the same time,the loss function is improved to reduce the loss weight of the background and the effect of positive and negative sample imbalance.In the longitudinal contrast,each step of the improvement has achieved varying degrees of profit.(4)In order to show the performance of the improved method in different scenarios and verify the effectiveness of the improved method for network transformer defect detection,different experimental tests are carried out.In attention module embedding experiment,the attention module is embedded in the classification network to experiment on the Image Net dataset,and then the attention module is embedded in the YOLO v3 module to experiment on the NTDD dataset,and results show that the proposed PAM and SAM module achieves the best optimization effect with the introduction of minimum parameters.In target and defect detection experiments,vertical comparison experiments on VEDAI and NTDD datasets are carried out to analyze each step to improve the optimization of the network,and lateral comparison experiments with Efficientdet-D1,YOLO v3 and YOLO v4 on COCO,VEDAI and NTDD datasets,and the experimental results show that the improved method achieves the best detection accuracy m AP 27.2%and 89.1%on VEDAI and NTDD datasets respectively.Meanwhile the precision of the method is further improved to 90.2%by pre-training strategy.The proposed method is evaluated in network transformer defect detection,and the results show that the total recognition accuracy is 97.3%,the total recall rate is 98.8%and the total missed detection rate is 0.2%when the confidence threshold is 0.5.Lastly,using Py QT5 to make user-product interaction simple and convenient GUI,in the implementation of the function to further verify the effectiveness of the method for network transformer defect detection.Experimental results show that the improved algorithm can improve the detection accuracy of small targets,meanwhile the internal defects of network transformer can be detected quickly and accurately.The results provide a theoretical reference for the realization of network transformer automatic visual inspection.