A fundamental framework for image matching based on large image features dataset

Abstract

Image matching is a fundamental component of various compelling applications such as object recognition, 3D-reconstruction, and image stitching. Matching images is basically applied through matching visual features. Scale Invariant Feature Transform (SIFT) [11], a state-of-the-art feature extraction approach, identifies distinctive feature keypoints in an image that are invariant to scale and orientation. The authors use the distance ratio (DR) test to match the SIFT features of two images. Although the popular DR test is simple, fast, and easily applicable to image feature matching, it is only based on heuristics. In this thesis, we aim to find a good match for a query image feature within a database of features. We apply SIFT on a huge ground truth dataset of image patches to extract the feature descriptors and compute the distances between the features. We use the Bayesian decision theory, which minimizes the probability of error, to achieve the optimal features' matching. However, the Bayesian approach is mathematically intractable in this case. Instead, we adopt a linear discriminant model (LDM) for classification. LD models, while suboptimal, are powerful classifiers and smoothly amenable to analysis. The LDM used, and especially the sorted LDM, one particular LDM classifier, is a generalization of the DR test. Leveraging the dataset of matching and non-matching distances, we thus analyse the DR test and the LDM models and investigate fitting the best model based on the structure of the images to be matched. Our analysis fundamentally justifies the capabilities of the DR test in identifying correct matches. We show that the DR test, by deciding based on the first and second minimum distance, is a particular case of the LDM classifier. However, since the DR test is conservative when classifying a feature match, we found that it misses true matches. When having a repetitive structure in the image, where the query feature matches more than one feature in the candidate matching image, the DR test ignores the