Surface defect detection of cylindrical lithium-ion battery by multiscale image augmentation and classification

While deep convolutional neural networks (CNNs) have recently made large advances in AI, the need of large datasets for deep CNN learning is still a barrier to many industrial applications where only limited data samples can be offered for system developments due to confidential issues. We thus propose an approach of multi-scale image augmentation and classification for training deep CNNs from a small dataset for surface defect detection on cylindrical lithium-ion batteries. In the proposed Lithium-ion battery Surface Defect Detection (LSDD) system, an augmented dataset of multi-scale patch samples generated from a small number of lithium-ion battery images is used in the learning process of a two-stage classification scheme that aims to differentiate defect image patches of lithium-ion batteries in the first stage and to identify specific defect types in the second stage. The LSDD approach is an efficient prototyping method of defect detection from limited training images for quick system evaluation and deployment. The experiments show that, based on only 26 source images, the proposed LSDD (i) constructs two augmented multi-scale datasets of 19,309 and 6889 image patches for training and test, respectively, (ii) achieves 93.67% accuracy for discriminating defect image patches in the first stage, and (iii) reaches 90.78% mean precision rate and 93.89% mean recall rate for defect type identification in the second stage. Our two-stage classification scheme has higher defect detection sensitivity than an intuitive one-stage classification scheme by 0.69%, and outperforms the one-stage scheme in identifying specific defect types. For comparing with YOLOv3 detector, less defect misdetections are observed in our approach as well.