Improved Faster R-CNN For Surface Defect Detection Of Flexible Printed Circuit Board Based On Decoupling Concept

Flexible printed circuit is a kind of printed circuit made by flexible substrate,and surface defect detection is one of the crucial steps to ensure the quality of flexible printed circuit board production.Although the method of manual visual inspection is gradually replaced by highspeed computer vision methods,computer vision detection methods based on traditional image features are affected by factors such as ambient illumination and image texture,and cannot handle images of flexible printed circuit board surface defects with complex backgrounds and textures.The deep learning object detection method represented by Faster R-CNN has better feature extraction and classification accuracy than the previous methods.However,the existing model is hard to handle the surface defects of flexible printed circuit boards with variable shapes and grant scale differences.Therefore,this paper addresses the characteristics of flexible printed circuit board surface defects and the problem of feature mismatch that exists in model coupling.The main work contents are as follows.(1)In order to address the characteristics of flexible circuit board surface defects with large-scale variation and random shape,as well as the problem of inconsistent feature information required by the RPN network and classification regression network,a hierarchical decoupled RPN target detection model is proposed.Firstly,the residual structure and multiscale feature fusion structure are used in the feature extraction network to enhance the detail and multi-scale feature information beneficial to locate and classify the surface defects of flexible circuit boards.Then the multiple receptive field fusion module is proposed,and a hierarchical decoupled RPN network model constructed by using several independent multiple receptive field fusion modules before the RPN network input to reduce the coupling with the classification regression network while increasing the receptive field size and high-level semantic information in the RPN network input feature map that is beneficial to defect localization.Finally,we determine the optimal parameters of different modules by orthogonal experiments.The m AP and recall of the final model are 0.9572 and 0.9696,which are better than other object detection methods.(2)The classification-network-guided method for detecting surface defects on flexible circuit boards is proposed for the problem of mismatch of required features between classification and regression tasks and poor classification performance in classification regression networks.Firstly,the classification decoupling module and the regression decoupling module are used to enhance the required feature information for the classification task and the regression task and realize the decoupling of the two tasks.Secondly,the SERes Net-18 model with the highest classification accuracy in the flexible printed circuit board surface defect classification dataset is proposed.Next,we propose a modified classification decoupling module that uses the specific defect category features extracted by the SERes Net-18 network to guide the feature extraction of the classification part of the classification regression network of the object detection model.The final m AP and average recall of the model are 0.9598 and 0.9697,which outperform other existing object detection models and prove the effectiveness of the model improvement.