A New Feature Descriptor for Multimodal Image Registration Using Phase Congruency.

Guorong Yu,Shuangming Zhao

doi:10.3390/s20185105

Abstract

Images captured by different sensors with different spectral bands cause non-linear intensity changes between image pairs. Classic feature descriptors cannot handle this problem and are prone to yielding unsatisfactory results. Inspired by the illumination and contrast invariant properties of phase congruency, here, we propose a new descriptor to tackle this problem. The proposed descriptor generation mainly involves three steps. (1) Images are convolved with a bank of log-Gabor filters with different scales and orientations. (2) A window of fixed size is selected and divided into several blocks for each keypoint, and an oriented magnitude histogram and the orientation of the minimum moment of a phase congruency-based histogram are calculated in each block. (3) These two histograms are normalized respectively and concatenated to form the proposed descriptor. Performance evaluation experiments on three datasets were carried out to validate the superiority of the proposed method. Experimental results indicated that the proposed descriptor outperformed most of the classic and state-of-art descriptors in terms of precision and recall within an acceptable computational time.

Highlights

Image registration is a vital step in many computer vision tasks, such as three-dimensional (3D)image reconstruction, image stitching, super resolution, and medical image processing
The second part of the proposed descriptor is based on the orientation of minimum moment of phase congruency
The minimized moment of phase congruency is the indication of feature points, which is assumed to be maintained in a multimodal image pair

Summary

Introduction

Image registration is a vital step in many computer vision tasks, such as three-dimensional (3D)image reconstruction, image stitching, super resolution, and medical image processing. Methods for image registration can be categorized into two types: feature-based methods and area-based methods [1,2]. Feature-based methods focus on the local information around the features (for example: lines, corners, and areas) and use this information to find the correspondence in the feature space. Area-based methods, called template matching, utilize a local window to find the correspondence according to some similarity metrics such as sum of squared difference (SSD), normalized cross-correlation (NCC), and mutual information (MI). Both of these types of traditional methods cannot be applied to multimodal image registration because of the non-linear tone mappings in multimodal images. Scale invariant feature transform (SIFT) [3,4] is widely used in image registration, as it can correctly describe the local information around the feature points

Methods

Results

Discussion

Conclusion