A Lightweight Multiscale Attention Semantic Segmentation Algorithm for Detecting Laser Welding Defects on Safety Vent of Power Battery

Abstract

At present, in order to improve the safety performance of power battery, a safety vent is welded on the battery cover to avoid unpredictable explosions. It is vital to detect the laser welding defects on safety vent effectively for product quality. In this paper, a lightweight multiscale attention semantic segmentation algorithm with high accuracy and efficiency was proposed. We built an experimental dataset of safety vent welding defects with a total of 7263 original images, which were collected from a battery manufacturing production line. The main framework of the proposed model consists of four modules: the improved Res2Net serving as the feature extraction sub-module, an attention mechanism, a localization block and a boundary anti-aliasing module. This architecture can segment defects of different sizes and shapes in real-time and get more refined segmentation results simultaneously. To evaluate the method, experiments concerning mean IOU and pixel accuracy were conducted, and an average validation accuracy of 99.4% and the mean IOU of 84.67% were achieved respectively. Furthermore, comparison experiments using some outstanding algorithms on safety vent’s welding defects test dataset were performed. It proves that our method achieved the best performance in terms of model size, computational complexity, efficiency and detection accuracy. Specifically, the model size is only 3.8 MB, and the frames per second (FPS) is 132.3. In brief, the proposed model is suitable for laser welding quality detection on safety vent in an industrial environment. Additionally, our study can provide a reference for designing relevant defect detection tasks using semantic segmentation method.