This paper introduces a novel semantic segmentation network, designed for object segmentation in power scenarios, and applicable to intelligent power inspection robots. The core technology is the DeepLab-MDA semantic segmentation model, an improvement on DeepLabv3+. This model integrates MobileNetv2, DenseSPASPP, IW, and AFF modules, effectively enhancing the segmentation accuracy of small target objects while maintaining real-time processing capabilities. On power-related datasets, the AFF module enhances the segmentation ability for small target objects such as power meter readings. The DenseSPASPP optimizes the depiction of object boundaries in power scenarios, such as oil stains and covers. The IW module improves the model’s generalization to environmental changes. The use of Focal Loss and Cos-logh Dice Loss addresses the issue of sample imbalance. Experimental results demonstrate the high performance of DeepLab-MDA on multiple datasets and its accuracy in practical deployment.