Surface Defect Detection For Lithium Battery Based On Multi-scale Features

With the continuous advancement of industrial technology,the industrial manufacturing process of lithium batteries has been basically achieved automation,but its surface defects detection still relies on low-efficiency poor-stability human inspecting approach.Using new technologies such as machine vision and deep learning to automatically inspect various battery surface defects can not only greatly improve product quality and production efficiency,but also promote the intelligent development and industrial evolution of the entire industrial field.Aiming at the several complex defects existing on the surface of soft-packed lithium batteries,an end-to-end defect detection method is designed using deep learning,and the algorithm was multi-dimensionally improved from data acquisition,feature extraction and analysis,and morphological segmentation,etc.First of all,considering the high-precision detection requirements of surface defects of lithium batteries in actual industry,the basic framework of target detection methods based on region proposal is described by using Faster R-CNN as an example.The core modules such as feature extraction,regional proposal networks(RPN),feature pooling method and multi-task loss function are analyzed in detail.Then,in order to solve detection problem caused by the excessive size span of the surface defects and the micro-defects of the battery,a multi-scale feature extraction and analysis method based on global view is proposed.On the basis of the feature pyramid network,the fusion of the macroscopic features in the small-scale feature map as well as the micro features in the large-scale feature map is strengthened by dense connection,and the feature denoising is performed by the deconvolution up sampling method to ensure the reliable connection.At the same time,the multiple hierarchical features are aggregated by the adaptive feature pooling to enhance the multiplexing of multi-scale features.Finally,based on the characteristics and detection requirements of lithium battery surface defects,an improved defect detection method based on region proposal is designed.Different data acquisition methods are discussed herein,and a multi-channel data construction method based on multi-lighting scheme is designed to achieve effective acquisition of depth information in pits,pinholes and other defects.Aiming at the evaluation requirements of the severity of defects in actual industrial detection,a lightweight defect segmentation network based on multi-stage sampling is designed to achieve accurate segmentation of defect contours.At the same time,Dice loss and binary cross entropy loss are combined to design an improved multi-task loss function,taking into account the convergence speed and stability.In the experimental verification,the extraction and fusion results of multi-scale features is analyzed firstly.To verify the effectiveness of the proposed method,the battery samples provided by the cooperative enterprise are used to systematically evaluate the detection accuracy,detection efficiency and segmentation effect of various methods.