Research On Surface Defect Detection Technology Of Automobile Panel Based On Domain Adaptation And Attention Mechanism Transfer Learning

Automotive cover parts are an important part of automobile body,and surface defects will seriously affect the subsequent forming and painting quality.Therefore,accurate and efficient detection of surface defects of automobile cover parts in production process is an important step to ensure product quality.With the development of machine vision and deep learning theory,surface defect detection based on deep learning has become a hot research topic for its high-precision real-time defect detection.However,due to the high yield of industrial production line,it is difficult to collect a large number of defective image data.At the same time,labeling the data is timeconsuming and laborious.The size of data set can not support the training of detection model based on deep learning,which ultimately leads to the insufficient performance of model detection.Therefore,this paper studies the surface defect detection method of automotive panel based on domain adaptation and attention mechanism transfer learning.The main research contents are as follows:(1)The transfer learning dataset and the overall framework of the defect detection model based on transfer learning are constructed.Aiming at the transfer learning problem of surface defects of automobile panels,the source domain aluminum profile surface defect dataset and the target domain automobile panel surface defect dataset are respectively constructed,and the defect images are preprocessed based on Z value normalization.The two-stage detection algorithm Faster R-CNN is selected as the basic detection model,and the overall framework of defect detection based on transfer learning is built.The network structure and parameters are adjusted accordingly based on the defect characteristics of the cover,and the experimental comparison and analysis of common detection frameworks are carried out.(2)A feature extraction method based on coordinate attention mechanism for surface defects of automobile panels is proposed.Aiming at the problem that the VGG network has insufficient ability to extract the surface defect features of the car panel,combined with the coordinate attention mechanism,a deep residual network based on the attention mechanism is proposed to extract the surface defect feature of the car panel.Some parameters of the network are determined through experiments.The effectiveness of the method was verified.(3)A defect domain adaptation method based on multi-scale features is proposed.To solve the problem of single-scale and foreground background feature alignment in global feature alignment of domain adaptive algorithm,multi-scale feature cross-layer fusion based on bilinear interpolation and defect image background suppression based on global average pooling are proposed,and a defect migration learning model for automobile panel surface based on multi-scale feature fusion domain adaptation is constructed,which realizes effective detection of surface defects of automobile panel under small sample conditions.(4)A surface defect detection system for automobile panel based on transfer learning was developed.The system includes data loading and marking module,model training and preservation module,surface defect detection module,and the system is used to effectively detect surface defects of automobile panel under small sample conditions.