Research On Defect Detection Method Of Resin Lens Based On YOLO

It is inevitable that there will be some defects in the production and processing of resin lenses,and it is necessary to detect the defects to control the product quality.At present,in the manufacturing industry,the detection methods of resin lens defects are still mainly manual detection,which can not meet the needs of automation development.The existing defect detection methods based on traditional machine learning and image processing have some problems such as poor generalization ability and slow detection speed,which are difficult to be applied in practice.Based on the above background,this dissertation studies a defect detection method of resin lens based on YOLO.The main work of this dissertation is as follows:Firstly,the types and shapes of resin lens defects are analyzed,and the target detection algorithm based on convolutional neural network is determined as the basic method of defect detection.Through the performance comparison experiment of target detection algorithm,combined with the real-time requirement of industrial detection,YOLOv5 s is selected as the basic algorithm of resin lens defect detection.Then,it focuses on the improvement of YOLOv5 s algorithm.Firstly,in order to reduce the pixel loss caused by the maximum pooling operation in the original algorithm,the hole convolution technique is used instead of the maximum pooling operation to form the pyramid pooling structure of the hole space,and this structure is added in front of the other two detectors to strengthen the fusion of local features and global features;Secondly,in order to enhance the expression ability of small defect features such as pits and bubbles,attention mechanism is added to the network,and attention modules are embedded in the residual blocks of the network to give weight parameters to the defect features;Thirdly,in order to improve the positioning accuracy of the prediction frame,the complete intersection ratio algorithm is used to replace the generalized intersection ratio algorithm and improve the loss function.The improved defect detection algorithm of resin lens is named YOLOv5s-AAM.Finally,the overall process of resin lens defect detection is determined and verified by experiments.The image of resin lens is collected in dark field,and then the image preprocessing and data enhancement are carried out to make the image data set of resin lens defect.Comparative experiments of detecting resin lens defects were carried out with YOLOv5 s,YOLOv5s-AAM and Faster R-CNN algorithms respectively.Experimental results show that YOLOv5s-AAM has the highest detection accuracy among the three algorithms,and can meet the real-time detection requirements.This method can be used for industrial real-time detection of resin lens defects and has strong generalization ability.