Adjacent Self-Similarity 3-D Convolution for Multimodal Image Registration

Significant challenges exist in the registration of multi-modal images (MMIs) due to nonlinear radiation differences, variations in lighting, and interference from image noise. These issues often lead to unreliable similarity measurements and low accuracy in point matching during multimodal registration. To address these challenges, this paper introduces a novel MMI registration method based on adjacent self-similarity three-dimensional convolution (ASTC). The proposed method consists of three main steps: feature point extraction, where key points are uniformly extracted via the block-FAST method; ASTC salient feature construction, where a local adjacent self-similarity model is employed to create multidimensional features; and feature structure enhancement, where a three-dimensional convolution is used for feature enhancement and finishing the process of image feature description. This paper evaluates the ASTC method against six sets of representative MMIs and compares it with six other algorithms. The results demonstrate that (1) the ASTC algorithm effectively overcomes radiation distortion, intensity differences, and lighting differences in MMIs, leading to improved accuracy in point matching. (2) The ASTC algorithm achieves higher matching efficiency and reduces time consumption, making it a practical choice for various data types. In summary, the proposed ASTC algorithm offers a robust solution for reliable registration of multimodal images, addressing common challenges related to image differences and improving the overall accuracy of the process. The experimental data and code link used in the paper can be found at https://github.com/yangwill81/ASTC.