A Lightweight Model for Real-Time High-Precision Leaky Cable Segmentation

In industrial inspection scenarios, precise localization of leaky cables is crucial for ensuring communication system stability. However, traditional manual detection methods often suffer from high misjudgment rates and poor real-time performance. To address the challenges of blurred boundaries, inadequate multi-scale feature fusion, and background interference in leaky cable image segmentation, this paper proposes an optimized DeepLabV3+ model. The proposed methodology incorporates several key enhancements: (1) applying Laplacian sharpening to enhance edge features; (2) introducing dense connections into the ASPP module to improve multi-scale feature propagation; (3) integrating the ECANet module to strengthen attention on critical feature channels; (4) replacing the original backbone with MobileNetV2 to improve computational efficiency; and (5) employing a hybrid cross-entropy and focal loss function to balance class weights during training. Experimental results demonstrate that the enhanced model achieves a mean Intersection over Union (mIoU) of 90.89%, a pixel accuracy of 98.78%, and a recall of 99.58%. With a processing speed of 196.4 frames per second (FPS), the model provides a robust and real-time solution for industrial leaky cable segmentation, effectively balancing high precision with operational efficiency.