The study of various registration methods based on maximal stable extremal region and machine learning

ABSTRACT

This review article aims to discuss contemporary and traditional image registration methods, which involve establishing correspondences between two or more images of a same object. Two popular methods based on local features are maximal stable extremal region (MSER) and deep learning (DL). After evaluating various registration methods, it has been determined that MSER and DL algorithms are well organised, convenient and highly accurate. These methods can effectively handle different aspects of images, such as size and viewpoint, and are currently in trend for image and video registration due to their excellent repeatability, efficiency and feature extraction capabilities. The article encompasses a comprehensive analysis of approximately 35 registration algorithms based on MSER and DL, spanning 2004–2023. MSER and DL techniques can be modified to specific requirements at different stages of the registration process. A comparative study was also conducted to assess the advantages and disadvantages of implementing each unique registration mechanism.

KEYWORDS:

• Deep learning
• maximal stable extremal region
• image registration
• interest point
• feature mining
• feature correspondence

Disclosure statement

No potential conflict of interest was reported by the author(s).