Reliable Image Matching Based On Hessian-Affine Detector and MROGH Descriptor

Reliable image matching is a vital step in many photogrammetric processes. Most image matching methods are based on the local feature algorithms because of their robustness to significant geometric and radiometric differences. A local feature is generally defined as a distinct structure with properties differing from its immediate neighbourhood. Generally, local-feature-based image matching methods consist of three main steps, including feature detection, feature description and feature correspondence. In the feature detection step, distinctive structures are extracted from images. In the feature description step, extracted features are represented with descriptors to characterize them. Finally in the correspondence step, the extracted features from two images are matched using particular similarity measures.
n this paper, an automatic image matching approach based on the affine invariant features is proposed for wide-baseline images with significant viewpoint differences. The proposed approach consists of three main steps. In the first step, well-known Hessian-affine feature detector is used to extract local affine invariant features in the image pair. In Hessian-affine detector a multi-scale representation and an iterative affine shape adaption are used to deal with significant viewpoint differences including large scale changes. To improve the Hessian-affine detector capability, an advanced strategy based on the well-known UR-SIFT (uniform robust scale invariant feature transform) algorithm is applied to extract effective, robust, reliable, and uniformly distributed elliptical local features. For this purpose, a selection strategy based on the stability and distinctiveness constraints is used in the full distribution of the location and the scale.
In the second step, a distinctive descriptor based on MROGH (Multisupport Region Order-Based Gradient Histogram) method, which is robust to significant geometrical distortions, is generated for each extracted feature. The main idea of the MROGH method is to pool rotation invariant local features based on intensity orders. Instead of assigning a main orientation to each feature, a locally rotation invariant schema is used. For this purpose a rotation invariant coordinate system is used to compute the pixels gradient. To compute descriptor, the pixels in the feature region are partitioned into several groups based on their intensity orders. Then, a specific histogram based on the pixels gradient magnitude and orientation is calculated for each group. Finally, the MROGH descriptor is generated by combining the values of all the gradient histogram from all groups into a single feature vector.
Finally, feature correspondence and blunder detection process is performed using epipolar geometry based on fundamental matrix. The initial matched features that are not consistent with the estimated fundamental matrix are identified as false matches and eliminated. A distance threshold TE = 1 pixel between each feature point and its epipolar line is considered as elimination condition. The experimental results using six close-range images show that the proposed method improves the matching performance compared with several state-of-the-art methods, including the MSER-SIFT, UR-SIFT and A-SIFT, in terms of the number of correct matched features, recall and positional accuracy. Based on the matching results, the proposed integrated method can be easily applied to a variety of photogrammetric and computer vision applications such as relative orientation, bundle adjustment, structure from motion and simultaneous localization and mapping (SLAM).