Super-Resolution Integrated Semantic Segmentation Method for the Corner Position of Catenary Bolt

Locating the corner position of a hexagon bolt with high precision is crucial, especially when dealing with low-resolution images during the tightening of wrist bolts in overhead contact systems. We present a novel framework for an SR-integrated semantic segmentation method. In this paper, we select the high-performance SRGAN model for low-resolution image reconstruction; then improve the Deeplabv3+ segmentation model to achieve efficient segmentation of hexagonal bolts; and finally determines the bolt corner positions based on Line Segment Detector. The novelty of this method lies in integrating image super-resolution into the semantic segmentation model, followed by the improvement of the segmentation network using the lightweight MobileNetv2 backbone, ultimately achieving precise bolt corner location. Experiments demonstrate that the proposed method improves corner detection accuracy by 38.93% compared to the original low-resolution method across different scenarios, proving its practical engineering significance.