Real-time cable defect detection algorithm based on MCN-YOLO

Abstract

To address the challenges posed by the high variety of cable defects, significant scale variations, and large aspect ratios, which make detection difficult, we propose MCN-YOLO, an improved model based on YOLO. First, an improved feature pyramid network, Multi-Scale Feature Fusion SPPF (MSF-SPPF), is introduced. By incorporating channel compression and depth-wise separable convolutions, the proposed module effectively reduces model parameters and computational complexity. Meanwhile, multi-scale adaptive pooling branches combined with residual connections enable efficient information interaction across different receptive fields, while channel attention further enhances sensitivity to small-scale defects. Second, the Convolutional Block Attention Module (CBAM) attention mechanism is incorporated into the neck network to highlight potential target regions and suppress background interference. Third, Normalized Wasserstein Distance (NWD) is adopted as the bounding box regression loss. By modeling bounding boxes as 2D Gaussian distributions and measuring similarity via Wasserstein distance, the model reduces sensitivity to minor positional deviations during training and improves localization robustness for small-scale targets. Finally, the model is converted and deployed onto the RK3588 embedded platform to achieve real-time cable defect detection. Experimental results demonstrate that the improved model enhances extraction capabilities for multi-scale defect features and small targets while maintaining a low computational load, effectively realizing real-time detection.