Research On PCB Bare Board Defect Classification Based On Few Shot Learning

With the development of PCB manufacturing industry,the demand for PCB defect detection becomes more and more complex.Therefore,defect detection technology based on image recognition is widely used in defect detection of various PCBs.However,in practical applications,the cost of manually marking defective images is often high,and a small number of defective samples can easily lead to over-fitting of the network.At the same time,the embedded module is not enough to extract the picture information,which reduces the accuracy of classification.This paper takes PCB defect images as the main research object.Aiming at the above-mentioned problems,it starts with the research from the perspectives of extracting features from the embedded module of few shot learning network and reconstructing the multiple loss function.The main research contents are as follows:1.This paper uses channel attention mechanism and fusion attention mechanism to extract important feature information in the embedded module.Perform feature extraction on the feature vector input from the upper layer,and use the residual module to map the feature vector input from the upper layer to the channel attention unit.Then introduce the fusion attention mechanism into the embedded module.The overall structural unit is connected in series.The combination of these two attention modules can not only effectively avoid the problem of feature value disappearance caused by the attention unit,but also strengthen the network's extraction of important features.2.This paper constructs a new loss function based on the original cross-entropy loss function of the few-shot learning network.Finds the feature map of the same category as the support set label from the feature vector map of the training set.Use the cross-entropy loss function for the same-category feature map as the intra-class loss,and use the cross-entropy loss function for the other different types of feature maps as the inter-class loss,and integrate the four inter-class losses to the same order of magnitude as the original loss function.This design can make the intra-class loss gradually decrease with the training iteration error,and the inter-class loss gradually increase with the training iteration,further improving the classification performance of the network.3.This paper combines the improved residual attention module and the fusion attention mechanism module embedded in the module with the reconstructed multiple loss function.This enables the embedding module to extract semantic and texture features of different scales,and further improves the ability of the few-sample learning network to extract features.In addition,the swish function is introduced into the embedded module,which effectively reduces the dependence between parameters and alleviates the occurrence of over-fitting.Better performance of the non-linear modeling ability of the few-sample learning network.The improved few shot learning method designed in this paper can accurately and efficiently identify 6 types of common PCB defects by improving the embedding module of feature extraction and reconstructing the loss function.For different basic models of learning network with few samples,this paper selects public data sets of different sizes,numbers,and complexity for testing.The experimental results show that the improved few shot learning network not only improves the accuracy of the network on the target data set,but also improves the adaptability of the network on the non-target data set.The method proposed in this paper can train a high-performance classification network suitable for image classification scenarios with a small amount of data.