Fine-Grained Hierarchical Perception Siamese Network for Battery Surface Defect Detection

Battery surface defect detection is an essential part of maintaining quality standards in battery production. However, challenges persist when addressing defects with complex and variable shapes, particularly when there is low separability between the defects and the background. To tackle these issues, this paper introduces the Fine-Grained Hierarchical Perception Siamese Network for Battery Surface Defect Detection (FHS-Net), which detects defects by leveraging the differences between the inspected image and a reference image. The framework comprises a weight-sharing visual encoder, a Fine-Grained Enhancement Module (FGEM), and a Hierarchical Perception Decoder (HPD). Specifically, FGEM fuses adjacent stage feature maps from the backbone to capture fine-grained details at shallow levels and semantics at deeper levels, improving defect feature extraction. HPD calculates loss at each decoding layer, forming a hierarchical perception that enhances defect awareness from intermediate layers, progressively improving recognition accuracy. Furthermore, to enrich defect detection datasets, we collected a battery surface defect dataset named BSD, which consists of 2,183 images obtained from actual production lines using an automated optical inspection (AOI) camera. Comprehensive experiments performed on the BSD, NEU-SEG, and Magnetic-Tile-Defect datasets highlight the outstanding performance of the proposed framework.