Enhancing Image Registration Leveraging SURF with Alpha Trimmed Spatial Relation Correspondence

In medical image registration, traditional approaches such as Harris corner detector, Oriented FAST and Rotated Brief (ORB), and scale invariant feature transform (SIFT) encounters several challenges for achieving precise alignment. These challenges arise from their limited ability to adapt scale, rotation, and illumination variations. To confront these challenges, the proposed methodology leverages the speeded-up robust features (SURF) algorithm for feature point extraction. Additionally, it employs the alpha trimmed spatial relation correspondence (ATSRC) algorithm as an efficient alternative to the conventional random sample consensus (RANSAC) method to effectively eliminate mismatches. In the pre-processing phase, the reference image undergoes Gaussian filtering to remove noise, and subsequent normalization is applied. Furthermore, feature detection is performed on both the reference and target images using SURF. A brute force (BF) matcher is applied, followed by alpha trimmed spatial relation correspondence to remove outliers. The process continues with computing the homography matrix and generating the registered image. Numerous experiments are conducted on monomodal and multimodal medical images acquired from the Kaggle dataset. The proposed methodology outperforms existing technique by significantly improving mutual information (MI) from 2.53 to 3.654, 2.257 to 3.17389 and cross correlation (CC) from 0.9219 to 0.95475.