Research On Defect Detection Method Of Small Sample Metal Sheet Based On Deep Learning

With 351 million smartphone market shipments in China in 2021,the smartphone chip shield case,as a special accessory,has been a challenge to detect defects intelligently due to its thin,light and reflective surface.It is usually necessary to first identify defects in the processing raw material strip for shielding shells to reduce equipment damage and material waste caused by raw material defects,and further realize the finished product inspection on the production line to guarantee the quality of products leaving the factory.At present,many scholars have carried out research and exploration around microdefect identification,small-sample learning and model construction,and achieved a lot of practical results.However,deep learning-based defect detection algorithms still face some problems during industrial applications.This thesis takes the production of raw material strip steel and cell phone chip shielding shell as the object of research,focusing on the detection of minor defects as well as model construction and optimization of three aspects.The design of nonlinear attenuation strategy,optimization of strip surface defect detection algorithm,establishment of shield shell surface defect data set,and construction of segmentation detection network have been completed to achieve efficient detection of shield shell surface defects.The specific research includes:1.A strip steel surface defect recognition method is proposed.Considering the accuracy and speed requirements of strip surface defect recognition by convolutional neural networks,combined with the problem of overfitting of network models easily caused by sample scarcity.Drop Path method is introduced to avoid overfitting of EfficientNetV2 network,and a nonlinear decay strategy is proposed to optimize the stacking multiplicity of modules in the network,reduce the parameter hierarchy,and improve the network recognition speed;data augmentation and migration learning methods are applied to improve sample diversity and model recognition accuracy.2.A method is given to construct a cell phone chip shield shell surface defect dataset.According to the demand of shield shell detection accuracy,a shield shell image acquisition platform is built to obtain high-quality shield shell surface defect images;for the problem that the distribution of tiny white mark defects on the shield shell surface is dense and messy,and manual labeling is very difficult,an OTSU-based defect labeling method is proposed to assist manual labeling of the acquired defect images;the large scale dithering Copy-Paste method is used to expand the processed A large-scale dithering Copy-Paste method is used to expand the processed data to obtain the shielding shell surface defect data set.3.A cell phone chip shield shell surface defect segmentation network is constructed.In order to solve the problem that factors such as tiny and different scales of defects on the shielding shell surface affect the detection speed and accuracy,the LSDANet network based on long and short connection paths and dual attention is proposed,adding long connection paths and shortcut branches to the basic codec semantic segmentation model to improve the feature extraction ability of the network for different scales of defects;designing the channel and spatial attention mechanisms separately to improve the tiny defect The channel and spatial attention mechanisms are designed separately to improve the segmentation performance and improve the defect detection accuracy.4.Built a software platform for shielding shell defect detection.Based on the LSDANet network model,the defect detection software including the detection module,data expansion module,image acquisition module,and training module is completed using PyTorch,OpenCV,and PyQt5 development tools to visualize the detection results for effect verification.