LRseg: An efficient railway region extraction method based on lightweight encoder and self-correcting decoder

Abstract

This paper proposes a lightweight and efficient railway region extraction model LRseg, which provides technical support for detecting foreign objects on the railway. LRseg consists of a lightweight encoder, self-correcting decoder, and segmentation head. The lightweight encoder is based on lightweight principles and consists of the LR-unit and Focus module. The self-correcting decoder is based on SC-FFM (Self-Correcting Feature Fusion Module) and cascade structure, efficiently utilizing the context and spatial information. The segmentation head introduces PPM (Pyramid Pooling Module) to extract multi-scale context information. At the same time, the segmentation performance of LRseg is enhanced by CIRKD (Cross-Image Relational Knowledge Distillation). The method has 0.78 M parameters and 1.47 G FLOPs (Floating Point of Operations), and the model size and memory demand are only 2.98 MB and 37.5 MB, respectively. On the self-built Railway-seg and public Railsem19 datasets, the method achieved mIoU (mean Intersection over Union) of 92.2% and 92.4%, respectively. When the input image size is 512×512, on Jetson TX2 and personal computer with Intel Core i9-12900, the method achieves 18 FPS and 94 FPS (Frames Per Second), respectively. Combining real-time and accuracy, the method is advantageous over the mainstream models, suitable for on-board devices with continuous work. The code will be released at https://github.com/CV-Altrai2023/LRseg.