Research On Fast Surface Defect Detection Algorithm Based On DenseYolo

Surface defect detection is an important part of industrial production and has a significant impact on the quality control of industrial products.Manual detection methods are inefficient,and traditional detection methods based on machine vision are limited by low robustness and high development costs.Detection methods based on deep learning have become the current research trend.Surface defects in practical applications have problems such as complex and changeable shapes,large scale changes,and low contrast,at the same time,they are restricted by noise interference and strict industrial production requirements,which make detection methods based on deep learning still have great challenges in terms of network detection speed and performance.In view of the above problems,this thesis proposes a fast surface defect detection algorithm based on DenseYolo.The main work is as follows:(1)For the existing deep learning-based surface defect detection backbone networks,which are mainly classified networks such as VGG and Resnet,there are problems of network redundancy and insufficient flexibility.Based on the advantages of DenseNet and YoloV3,this thesis proposes to design a new type of DenseYolo network from the perspective of receptive field,improve the multi-layer prediction network of YoloV3 and design a Dense module to make the network lightweight and able to extract features of various scales.And improve the detection ability and positioning accuracy of low contrast defects.(2)In order to solve the problem of template matching method based on gray normalized cross-correlation,which needs to create a large number of multi-scale,multi-angle offline templates,this thesis has improved this.The area of the rotation template in the image to be matched is instantly obtained through the affine transformation matrix,and online arbitrary angle matching is achieved,which greatly reduces the program's memory requirements.At the same time,the pyramid image structure and the progressive matching strategy are used to greatly reduce the matching time.Compared with the general method based on image structure,this method is more robust and can overcome the variation of illumination and certain noise interference.(3)For the problem that the defect images collected on the production line cannot be directly used for network training and testing,this thesis first uses the proposed multi-angle template based on gray normalized cross correlation to locate the detection area,then uses the Yolo_mark tool to mark the type and range of defects,and divide them into training and test sets according to the proportion.In addition,the number of images in the training set is amplified based on data augmentation techniques.The surface defect detection algorithm proposed in this thesis takes the lithium battery surface scratches,lithium battery surface dents and pits dataset provided by the partner company Huizhou Gaoshi Technology and the open source steel surface defect dataset as verification targets.Experimental results show that the algorithm has good detection effect and detection speed,both of which are superior to Faster-Rcnn,SDD and YoloV3 networks.With the help of transfer learning technology,fine-tuning the trained network and then quickly applies it to new defect detection projects,which has good landing advantages and prospects.