Enhanced phase congruency feature-based image registration for multimodal remote sensing imagery

Abstract

Multimodal image registration is an essential image processing task in remote sensing. Basically, multimodal image registration searches for optimal alignment between images captured by different sensors for the same scene to provide better visualization and more informative images. Manual image registration is a tedious task and requires more effort, hence developing an automated image registration is very crucial to provide a faster and reliable solution. However, image registration faces many challenges from the nature of remote sensing image, the environment, and the technical shortcoming of the current methods that cause three issues, namely intensive processing power, local intensity variation, and rotational distortion. Since not all image details are significant, relying on the salient features will be more efficient in terms of processing power. Thus, the feature-based registration method was adopted as an efficient method to avoid intensive processing. The proposed method resolves rotation distortion issue using Oriented FAST and Rotated BRIEF (ORB) to produce invariant rotation features. However, since it is not intensity invariant, it cannot support multimodal data. To overcome the intensity variations issue, Phase Congruence (PC) was integrated with ORB to introduce ORB-PC feature extraction to generate feature invariance to rotation distortion and local intensity variation. However, the solution is not complete since the ORB-PC matching rate is below the expectation. Enhanced ORB-PC was proposed to solve the matching issue by modifying the feature descriptor. While better feature matches were achieved, a high number of outliers from multimodal data makes the common outlier removal methods unsuccessful. Therefore, the Normalized Barycentric Coordinate System (NBCS) outlier removal was utilized to find precise matches even with a high number of outliers. The experiments were conducted to verify the registration qualitatively and quantitatively. The qualitative experiment shows the proposed method has a broader and better features distribution, while the quantitative evaluation indicates improved performance in terms of registration accuracy by 18% compared to the related works.