Research on Power Line Semantic Segmentation Algorithm Based on Improved DeepLabv3+

The power transmission network is an important part of the power system. Intelligent robots that can connect and disconnect power lines while live can ensure the uninterrupted supply of electricity. Accurately segmenting the power lines from camera images is crucial for live-line work by robots. This task often faces challenges such as the power lines being too thin, complex and variable backgrounds in captured images, and the need for real-time segmentation of power lines. To address these challenges, this paper proposes a power line semantic segmentation algorithm based on an improved DeepLabv3+. The algorithm replaces the original backbone network with a lightweight MobileNetv4 to reduce the computational load of the model. It introduces the SimAM attention mechanism module to adaptively enhance the edge features of the power lines, improving the precision of feature extraction. Additionally, it combines a weighted cross-entropy loss with a Dice loss function to optimize segmentation tasks for small target areas. Experimental results on a standard dataset show that the proposed algorithm improves the F1-Score and mIoU by 1.97% and 1.54%, respectively, compared to the original algorithm.