Abstract
Large oil-immersed transformers have metal-enclosed shells, making it difficult to visually inspect the internal insulation condition. Visual inspection of internal defects is carried out using a self-developed micro-robot in this work. Carbon trace is the main visual characteristic of internal insulation defects. The characteristics of carbon traces, such as multiple sizes, diverse morphologies, and irregular edges, pose severe challenges for segmentation accuracy and inference speed. In this paper, a feasible real-time network (deformable-spatial-Canny segmentation network, DSC-SeNet) was designed for carbon trace segmentation. To improve inference speed, a lightweight unilateral feature extraction framework is constructed based on a shallow feature sharing mechanism, which is designed to provide feature input for both semantic path and spatial path. Meanwhile, the segmentation model is improved in two aspects for better segmentation accuracy. For one aspect, to better perceive diverse morphology and edge features of carbon trace, three measures, including deformable convolution (DFC), Canny edge operator, and spatial feature refinement module (SFRM), were adopted for feature perception, enhancement, and aggregation, respectively. For the other aspect, to improve the fusion of semantic features and spatial features, coordinate attention feature aggregation (CAFA) is designed to reduce feature aggregation loss. Experimental results showed that the proposed DSC-SeNet outperformed state-of-the-art models with a good balance between segmentation accuracy and inference speed. For a 512 × 512 input, it achieved 84.7% mIoU, which is 6.4 percentage points higher than that of the baseline short-term dense convolution network (STDC), with a speed of 94.3 FPS on an NVIDIA GTX 2050Ti. This study provides technical support for real-time segmentation of carbon traces and transformer insulation assessment.