Detection of Battery Appearance Defects Based on Multi‑Scale Object Detection

[Purposes] Based on YOLO v4 framework, a new feature extraction network VoVNet-A algorithm was designed, which can effectively identify image fine-grained features. [Methods] With the proposed algorthm, acts of improved attention module CSAM on aggregated features, distinguishes of the importance of different channels and regions of aggregated features, and efficient filterings of the redundant features brought by feature aggregation were realized. The selection of pre-selected boxes was also optimized and a variety of data enhancements were used to expand the defect data, which finally improved the detection of battery defects. [Results] The ablation experiments show that the above enhancements can improve the detection accuracy to different degrees. The comparison experiments show that compared with the commonly used target detection algorithms Fast RCNN, SSD-VGG16, and YOLO v4, the mAP values of the method for battery defects are improved by 11.5%, 21.5%, and 3.3%, respectively, and the FPS quantities are increased by 16, 12, and 4 frames, respectively.