Deep Learning-Based Image Registration for Train Chassis

Due to the complex features of train chassis images, conventional registration methods perform poorly. To improve registration accuracy and efficiency, this paper proposes a Transformer-based approach incorporating frequency-domain processing. Specifically, cosine similarity between template and deformed chassis images is computed first, serving as an indicator of deformation magnitude. Features are then extracted using an enhanced backbone network, with the cosine similarity integrated into these features to boost registration precision. For efficiency improvement, a neural network predicts the low-dimensional Fourier-domain representation of image pairs. A parameter-free decoder subsequently decodes this representation into the final full-resolution deformation field, accelerating inference. Experiments confirm the proposed method surpasses existing approaches in both accuracy and efficiency for train chassis image registration.