Performance Evaluation of Surface Defect Segmentation with a new Battery Surface Gray Dataset

Batteries are essential products for providing power to industrial equipment and devices. Factories producing batteries need to conduct defect segmentation on thousands of batteries, which is costly and inefficient when done manually. Early methods for defect identification, relying on features like pixel intensity, color distribution, or texture, often suffer from low accuracy and poor generalization, challenging adaptation to new data distributions. Industrial defect segmentation has gradually shifted towards using deep learning methods, which effectively address issues like generalization. However, a critical challenge remains: the lack of datasets. Existing datasets for battery defects are insufficient to support deep learning-based segmentation methods. In this paper, we proposed a new Battery Surface Gray Dataset (BSG Dataset), specifically designed for battery manufacturers to address defect segmentation tasks. Our dataset, collected from multiple production lines using specialized AOI cameras, encompasses diverse factors like illumination and temperature, leading to variations such as irregular placement and uneven lighting. It comprises 3546 high-definition battery images, each meticulously annotated with accurate labels for two defect classes: scratch and dent. Finally, we evaluated and compared the segmentation performance of 7 advanced defect segmentation models on the BSG Dataset.