Image Affine Invariant Feature Extraction And Matching

Feature extraction is a basic problem in image processing and computer vision applications, such as target recognition, image registration, image retrieval, and scene matching, ect. Since the images taken from different sensors and viewpoints usually suffer from distortions, the features should be capable of capturing the essence of targets and be less sensitive to the imaging. Affine invariant features has been widely researched due to its invariance to sensor parameter variance and viewpoint changing. However, it is still a challenging filed as a high quality feature is required to be robust, distinction, repeatability and application. It is discovered that human prefer to recognizing the target based on the shape rather than texture or color, thus, shapes are important information of targets. This thesis is primarily concerned with the problem of using new theory to extract affine invariant features from point sets, contours, regions, and relative distribution between shapes.As the foundation of this thesis, an affine transformation, which is used to approximate the geometric geometric relationship between images of different sensors and viewpoints, is exploried in Chapter 2 based on a pin-hole camera model. Meanwhile, according to matrix decomposition, the affine transform can be treated as the integration of three transformations including rotation, scaling and shearing, and the effect of each transform on the shape distortion is illustrated. Finally, some important characteristics of affine transformations are proved.In Chapter 3, the shape is considered to be a point set, and the problem of affine invariant shape feature extraction is converted to extracting features from point set. Based on the affine invariance of point set dividsion, a new affine invariant feature of point set is proposed using iterative dividing. The point set is divided into two parts in each division and the centroids of divided parts are extracted, the location of centroids obtained from all divisions in the from of binary tree results in a centroid tree(CT). Owing to the affine invariance, CT can be considered as a support point set, from which the features are further extracted. Coupled with the invariance of triangle area ratio and moment, two affine invariant features, CT-SPS-FSD and CT-SPS-DOPM are developed. The capability of the two features for point matching under affine transformations and noise are well tested.Depending on the affine invariance of shape projection, chapter 4 proposes two affine invariant features taking advantage of the contour projection support point set(CP-SPS) and the shape projection distribution(SPD) respectively. The two features are both obtained based on shape projection. However, different from CP-SPS which combines the support point set composed of the contour points with the sampled projection value and moment computation, SPD treats the histogram of the shape projection distribution as the feature. The advantage of two features for shape describtion is well demonstrated by experiments on synthetic and real data. However, the different information captured by two features results in their different applications. CP-SPS, which is sensitive to the local change of the contour, is more capable of similar contour discriminating. Whereas, SPD is more robust to noise since all information of points in the shape region are adopted and the relaxation labeling technique is introduced for point matching.In Chapter 5, a new proposed shape descriptor, named as ratio histograms(R-histogram), is constructed by the relative attitude relationship between two independent shapes. For a pair of two shapes, the shapes are treated as the longitudinal segments parallel to the line connecting centroids of the two shapes, and its R-histogram is composed of the length ratios of collinear longitudinal segments. Thus, R-histogram is theoretically affine invariant due to collinear distance invariance of the affine transformation. In addition, as the computation of the length ratio weakens the noise contribution, R-histogram is robust to noise. The shape matching algorithm based on R-histogram includes two major phases: preprocessing phase and matching phase. The first phase, which can be processed off line, are trying to obtain the R-histograms of all original shape pairs. In the second phase, for each transformed shape pair, its R-histogram is computed and the candidate matched shape pair with minimal R-histogram matching error is found. Then, a voting strategy, which makes the accuracy for shape matching further improved, is adopted on the candidate corresponding shape pairs. Experimental results demonstrate the proposed R-histogram is robust and efficient.Finally, Chapter 6 concludes the thesis and recommends future research directions.