New Hierarchical Approaches in Real-Time Robust Image Feature Detection and Matching

Abstract

The detection of robust image features of high distinctiveness forms a ubiquitous problem in digital image processing. Robust image features are the first step and the key to reliably detect patterns or objects in images, which subsequently leads to object classification and semantic interpretation of the objects in a given image. The ideal situation for object recognition and classification is to find unambiguous traits of patterns or objects that are to be processed. For anyone delving into this matter, the subject presents itself as a challenging task. Besides the often not obvious object traits, there are a lot of other issues that have to be taken into account. For instance, lighting conditions can vary as well as the objects’ scales and rotations; image noise or partial occlusion also is likely to occur. Unless one conducts experiments under laboratory conditions, where some of these problems might be ruled out, all these issues have to be addressed properly. So the challenging task is to find image features that are distinct and robust under the varying conditions just stated before. An obvious method for detecting objects is to examine their shape and, as an abstraction of it, their contours. Contour matching usually works well, if the distinct object classes have strong variations in their shape (like the shape of any automobile is quite different compared to the shape of a human). The contour can be represented by a classic chain-code (or its many derivates). For the matching process both contour representations undergo a transformation to achieve scale and rotation invariance and are then compared to each other. The comparison can either be done directly on the transformed image coordinates or on measures deduced from the earlier representation (i.e. the moments, distances, polygonal, or Fourier descriptors yielded from the contour points). Besides those methods, one popular approach is to detect certain local interest points at distinctive locations in the image, such as corners, blobs, or T-junctions. The interest point detector is supposed to be repeatable under varying viewing conditions (e.g. different lighting or different viewing angles). Having found distinctive interest points in an image, the interest points are examined more closely regarding their neighbourhood. Taking the neighboring pixels into account forms an image feature or image descriptor, which has to fulfill certain criteria regarding distinctiveness, robustness against noise, detection errors, and image deformations. These image features can then be matched to features computed from other images. Matching features indicate a high likeliness of correspondence between the two images. Hence patterns or objects in one image can be detected in other images O pe n A cc es s D at ab as e w w w .ite ch on lin e. co m