ArcMask: a robust and fast image-based method for high-speed railway pantograph-catenary arcing instance segmentation

Wei Quan,Shaopeng Guo,Xuemin Lu,Guoxin Gu,Wenjing Cheng,Ning Zhou,Long Yu
DOI: https://doi.org/10.1007/s00521-022-08059-7
2022-12-04
Neural Computing and Applications
Abstract:The pantograph-catenary arcing reflects the health of pantograph-catenary and current collection quality of high-speed railway, so the arc detection is of great significance. However, due to the scene complexity, intra-class polymorphism and inter-class similarity of arcing and the fast running speed of high-speed railway, it is still a huge challenge to achieve fine and robust arcing detection. To overcome these issues, a robust and fast image-based instance segmentation method called ArcMask is proposed to detect pantograph-catenary arcing, which designs a new attention-based multi-scale feature fusion module that combines both top-down and down-up modules to realize arcing pixel-level instance segmentation. The effective combination of instance-level information and bottom-level semantic information balances features representation ability of top-level and bottom-level features. Compared with other instance segmentation methods (e.g., BlendMask), it can effectively learn feature representation with tiny, irregular and complex arc features and speeds up the calculation. In addition, both deformable convolution and depth-wise separable convolution are introduced in ArcMask, which aims to improve the segmentation performance of irregular arcing and efficiency. The ArcMask can distinguish different arcing instances at pixel-level with fine granularity and distinguish inter-class and intra-class features of arcing, instead of just focusing on rectangular bounding box. Experiments on self-collected dataset IVAIS-PCA2021 verify the effectiveness and efficiencies of the ArcMask. Its AP, AP 50 and AP 75 are 56.61, 94.14 and 64.56, respectively, and the fastest reasoning speed based on MobileNet is 56 FPS. Compared with other state-of-the-art segmentation methods, the proposed ArcMask has better integrity in arcing edge detection.
computer science, artificial intelligence
What problem does this paper attempt to address?