YOLO-LCGE: a real-time instance segmentation model for non-destructive testing of defects with extreme scale variations on copper cathode plate surfaces

Cathode copper requires rigorous quality control through non-destructive testing methods. Extreme scale variations of defects on it pose significant challenges for existing models. An instance segmentation model named YOLO-LCGE is proposed in this paper, with three key innovations: a Context-Guided Feature Pyramid Network generating dual-path complementary features via channel importance-based adaptive fusion; a Multi-Scale Hybrid Convolution Segmentation Head constructing multi-frequency responses and continuous receptive field gradient and a Lightweight Cross-Stage Partial Structure reducing computational complexity through linear transformation. Experiments on the self-built dataset show a 7.0% improvement in mAP50-95 over the baseline with a speed of 177.6 FPS. Experiments on two public datasets demonstrate that YOLO-LCGE has a strong generalisation ability. The model is further deployed on an NVIDIA Jetson Orin NX edge computing platform, where it achieves an end-to-end inference speed of 28.6 FPS in FP16 precision, confirming its practical value for on-line non-destructive inspection of cathode copper plates. The work provides an effective solution for multi-scale industrial defect segmentation.