Research On Bend Surface Defect Detection Based On Deep Learning

Bend is a widely used workpiece in industry.It is inevitable that there will be damages or defects in the actual production of bend.The existence of bend defects will have a negative impact on the quality of products.Therefore,defect detection is an indispensable part in the industrial production of bend.At the same time,automatic production also puts forward higher requirements for the speed and accuracy of defect detection.Therefore,this thesis constructs a lightweight object detection algorithm based on deep learning to realize the rapid location and detection of the surface defects of the copper bend.(1)High-precision one-stage object detection network has complex structure,many parameters and high configuration required for training.In order to solve the above problems,a lightweight object detection algorithm,MA-YOLOv4(Mobilenetv2 with Attention-YOLOv4),was proposed.First,the backbone network of YOLOv4 was replaced with Mobilenetv2,and the network was improved through double residual convolution blocks and depthwise separable convolution,while greatly reducing the amount of network parameters;Then,to further improve the detection accuracy of the algorithm,CA(Coordinate Attention)algorithm and ASFF(Adaptively Spatial Feature Fusion)algorithm was used in MA-YOLOv4;Finally,in the inference stage,a reparameterization design was added to the double residual convolution block,which reduced the amount of model parameters and greatly improved the inference speed without losing detection accuracy.The algorithm’s universality and detection performance have been verified on the PASCAL VOC public dataset.Experimental results show that MA-YOLOv4 maintains high detection accuracy while being lightweight,which can meet the needs of mobile or small device object detection tasks.(2)To suppress overfitting and improve the generalization ability of the model,enrich the image data of bend defects and add data enhancement algorithms.The image is preprocessed offline,outliers are removed,and some defects are transplanted through local clipping and deformation to enrich the training data;At the same time,Mosaic,Mixup and other data enhancement algorithms are used to complete real-time data enhancement during the training process.(3)In order to achieve rapid and accurate detection of bend surface defects,the MAYOLOv4 algorithm was applied to the bend surface defect detection task and the performance of the network in the bend surface defect detection task was analyzed through experimental data.Focal Loss is added to the original MA-YOLOv4 to solve the imbalance between positive and negative samples of bend defect data,and K-means clustering algorithm is used to redraw the preset anchor frame.The experimental results show that after using image preprocessing,data enhancement and other methods,the detection accuracy of pipe surface defects has been improved to a certain extent.MAYOLOv4 shows good detection performance in defect detection tasks,providing an experimental basis for defect detection in automatic production.