Research On Industrial Product Appearance Defect Detection Technology Based On Feature Difference Spatial Vision Algorithm

In the current industrial production,defective products have gradually become the bottleneck for enterprises to improve product quality.The importance of industrial defect detection is self-evident.In industrial product appearance defect detection,machine vision detection has the advantages of high precision,fast speed and good stability compared with artificial vision detection.Therefore,with the help of deep learning technology,machine vision has been widely used in industrial defect detection.In addition,in industrial defect detection,due to the inevitable emergence of unknown defects,the supervised target detection algorithm based on data driven has been unable to be applied.This research is committed to studying the detection algorithm of unknown defects in industrial products,and proposes a zero sample industrial defect anomaly detection model Zero DD(Zero sample Defect Detection).By learning the characteristics of known samples,the model can still recognize unknown defect samples as defects when encountering them.The three important components of the model can ensure its sensitivity to unknown defects and achieve high-precision detection,which is also the tasks of this thesis.The main work of this thesis is as follows:First,a feature difference module is proposed,which is composed of a parallel architecture with shared weights.Based on the twin network structure,the module uses dual input channels to strengthen feature differences and weaken the same kind of attributes.The feature difference module can better highlight the defect features and effectively improve the detection accuracy of the model for small targets.Second,for the detection of unknown defects in industrial products,the difference in data distribution between intact samples and defect samples is used to couple the difference features with the input of the generator and feed them into the Generic Adversary Network(GAN).The research shows that GAN model can distinguish normal and defective samples by learning from known samples.When the model encounters unknown defects,it can not generate images that meet the discriminator to recognize the unknown defects.Third,this study proposes channel and coordinate attention modules to enable the network to focus more on defect features and further improve the model detection accuracy.This module adds channel attention on the basis of coordinate attention,and makes full use of attention mechanism to improve model performance.Fourthly,a new visual data set BC defects(Bottom Cap defects)for injection molded bottle cap defects was released.The data set contains 8 kinds of defects,with a total of 3008 samples.The total number of samples and the number of categories are more than other published industrial product data sets,and can support multi label classification.This data set can further improve the benchmark of industrial defect data set.Finally,the simulation experiments are carried out based on the proposed visual data set of injection bottle cap defects.The experimental results show that the proposed model is effective in industrial product appearance defect detection.It is worth mentioning that we regard different defects as unknown defects to verify the detection capability of the proposed model for unknown defects.