A Robust Image Registration Using Multi-Scale Feature Extraction Techniques Using Deep Learning

Image registration is a significant assignment in computer visualization and medical imaging, which involves aligning two or more descriptions of the same division or object taken from different viewpoints or at dissimilar epochs. Deep learning-based approaches have recently demonstrated promising outcomes in several image-processing applications, including image registration. Using multi-scale feature extraction methods and deep learning, the study provides a reliable image registration method in this study. Three primary steps make up the suggested method: feature extraction, feature matching, and transformation estimation. A deep convolutional neural network (CNN) is utilized in the feature extraction process to extract multi-scale information from the input images. The CNN architecture is built to manage large-scale deformations by extracting features at various scales. Following feature extraction, a strong matching technique based on the nearest neighbor distance ratio test is utilized to match the extracted features. The proposed method uses the Random Sample Consensus (RANSAC) algorithm to estimate the transformation parameters robustly, which helps to handle outliers and noise in the feature-matching process. The proposed method is assessed on several benchmark datasets, including the ETH-80 dataset and the TUM RGB-D dataset, and compared with state-of-the-art approaches. The investigational consequences demonstrate that the proposed method outstrips prevailing approaches in terms of accuracy and robustness, especially for images with large-scale deformations. The proposed technique also demonstrates good performance in the presence of noise and outliers, which further demonstrates its robustness.