Structure-Aware Deep Image Alignment for Automatic Wayside Rail Freight Vision

Abstract

The wayside rail freight machine vision system is widely used to facilitate the inspection and condition monitoring of freight trains, supporting safe operations and efficient maintenance. Within this system, image alignment or registration plays a crucial role, serving as a precursor to various downstream functions such as change detection, state tracking, and defect inspection. However, this task is severely challenged by the demanding 3L conditions—low texture, low illumination, and large displacement—which are prevalent in real-world inspection scenarios and cause significant performance degradation in existing alignment methods. To address these challenges, this paper introduces a novel structure-aware learning paradigm for deep image alignment. The core innovation is to explicitly guide the network to prioritize robust structural features over unreliable texture or intensity information. Specifically, we propose a synergistic combination of a structure-aware loss function and a structure learning module; the former directs the loss computation toward structurally salient regions, while the latter regularizes the backbone network to inherently comprehend image geometry. These modules are designed to be architecture-agnostic and work synergistically to embed structural knowledge into the alignment process. We instantiate this paradigm within two foundational deep alignment architectures. Experimental results on a challenging freight train dataset demonstrate that our approach not only establishes a new state-of-the-art in alignment accuracy under 3L conditions but also maintains highly competitive efficiency, providing a robust and practical solution for enhancing the reliability of wayside rail freight inspection and maintenance.