HOG descriptor based registration (A new image registration technique)

Abstract

This paper proposes a novel feature based technique that can be used to address image registration problems in the field of Digital Image Processing. The two major parts of this proposed method is key point description and the matching of those key points. This process requires two images, one as a reference and another as a target image, and applies FAST (Features from Accelerated Segment Test) Feature Detector to detect the key points in both images. For every detected key point, a rectangular region is selected and feature vector (feature descriptor) for respective key point is generated using Histogram of Oriented gradients (HOG). All the feature vectors are then matched in both the images to generate the Transformation Matrix (containing transformation values of image in x-y direction) which is used to align target image. An additional refinement stage is introduced to eliminate outliers and to improve registration accuracy. The proposed method has been tested on a number of general and microscopic images, and the method proves to be registering images with high accuracy in presence of different transformations.