CSD-YOLO: An improved lightweight defect detection network for capacitor side-surface defects

Capacitor surface defects, such as top bulges and breakage, significantly impact product quality and safety in industrial applications. To address the limitations of existing lightweight models in balancing detection accuracy and real-time performance, this paper proposes CSD-YOLO, an improved YOLOv7-tiny-based network. First, the GhostNet backbone reduces computational costs and parameters while maintaining feature representation. Second, deformable convolution (DCNv2) integrated into the ELAN-D2 module enhances irregular defect detection by adaptive receptive fields. Third, channel-wise knowledge distillation transfers knowledge from a YOLOv7 teacher model to the lightweight student model, improving accuracy without increasing complexity. Experiments on a custom dataset show that CSD-YOLO achieves a [email protected] of 98.9% with 0.75ms inference latency, outperforming baseline models in both efficiency and precision. The proposed model demonstrates robust performance for real-time deployment on resource-constrained devices.