Research on Infrared and Visible Image Registration Algorithm for Complex Road Scenes

This study proposes a novel image registration algorithm to solve the problem of low registration accuracy caused by excessive difference in resolution and spectral differences between infrared and visible images. First, we use a VI-CycleGAN network to translate visible images into fake infrared images. Furthermore, we incorporate a normalization-based attention mechanism (NAM) into each residual block to capture the global information of the image and preserve its details. Meanwhile, we consider higher-level semantic information and introduce a hybrid loss function that better preserves the content features of the original image. We then use guided filtering to process the infrared image and fake infrared image to eliminate complex background noise. Subsequently, coarse registration was performed using the Speeded-Up Robust Features (SURF) algorithm. Finally, we used the random sample consensus (RANSAC) algorithm to remove the mismatch points and achieve accurate registration. This study conducted experiments on two different VIS-IR image datasets and compared Gaussian field estimator with manifold regularization (GFEMR), radiation-variation insensitive feature transform (RIFT), and locally normalized image feature transform (LNIFT) algorithms. The experimental results show that compared with LNIFT algorithms, the registration accuracy of the proposed method was improved by approximately 5%.