Intelligent Pixel-Level Rail Running Band Detection Based on Deep Learning

Abstract

The rail running band is a mathematical representation describing the continuous strip-shaped spatial surface resulting from the rolling contact operation of train wheels on the rail surface, which establishes a direct mapping relationship with the wheel–rail interaction, and the nature of this interaction significantly influences the safety and comfort of train operations. Therefore, accurate detection of the running band is crucial for enhancing the safety and comfort of train travel. Traditional running band detection relies on manual inspection methods, utilizing a scale for measurements on the rail. However, this approach is characterized by high labor costs, slow detection speeds, and a lack of systematic data preservation. This paper proposes R2Bnet, a lightweight semantic segmentation algorithm that achieves pixel-level detection of rail running bands. R2Bnet is an enhanced encoder-decoder architecture built upon ShuttleNet. Different from ShuttleNet, R2Bnet optimizes the number of repetitive codecs in ShuttleNet and redesigns the encoder’s residual structure to match the unique characteristics of rail running bands, allowing the backbone network to effectively capture long-range dependencies. Furthermore, R2Bnet integrates an efficient channel attention mechanism to enhance focus on critical regions and optimize feature representations. The F-measure and mean intersection over union (mIOU) achieved by R2Bnet on 300 testing images were 98.47% and 0.9617, respectively. Notably, R2Bnet outperformed six state-of-the-art models for semantic segmentation and demonstrated a significant 39% improvement in speed compared with the average speed of the six networks provided.