Study Of The Construction Of Image Local Descripor

Local descriptor is a low-level distinctive and robust feature description, widely applied in many computer vision tasks such as image retrieval, image classification, object detection and recognition owing to its discrimination in image deformation, occlusion and cutting and robustness to various image transformation.Increasing applications drive the researchers to propose a variety of local feature descriptors with different advantages, and an ideal local descriptor is expected with high-quality description power, distinctiveness, robustness and fast construction and matching. But it is still a challenge to construct perfect local descriptor. This dissertation pays more attention to the construction of compact local descriptors and the local descriptors robust to illumination, rotation and noise.The main contents are summarized as follows:A compact local descriptor is presented, which is based on a terse texture feature called the dominant centre-symmetric local binary pattern (DCSLBP). DCSLBP has similar distinctiveness and half dimension compared against original centre-symmetric local binary pattern (CS-LBP). On the basis of an improved partition for image region, DCSLBP histogram is computed on every subregion under the coordinate system. Then the rotation-invariant histograms are concatenated according to the intensity order of subregions to construct the local descriptor. The local descriptor is extended to two variants with multi-scale region support and color space support. These descriptors are compared with scale-invariant feature transform (SIFT), multi support region rotation and intensity monotonic invariant descriptor (MRRID), orthometric combination local binary pattern (OC-LBP) in interest region matching and in the application of object recognition. The experiments demonstrate the proposed descriptor’s compactness and robustness to various image transformations, especially to large illumination change.Several short local descriptors with different dimensions are presented, which are based on the selected optimal pattern subset and the existing classic construction. The selection is conducted in non-greed way on large image patches set to achieve the optimal binary patterns subset. To avoid tremendous computation in training, a searching tree with Branch & Bound and simple prediction mechanism are used. The effectiveness of patterns selecting has be confirmed by the evaluation of these new local descriptors’ performance based on the selected pattern subset. These new local descriptors in MRRID-like construction are compact, distinctive and robust to image transforms, and the local descriptors in SIFT-like construction is more robust to noise disturbing.A fast and robust local binary descriptor using polar location, called rotation-invariant local difference binary descriptor (RLBD), is presented. In the partition of image region with multiple log-polar location griddings, binary tests of average intensity and the differences in the radial direction and in the tangent direction of sector region are computed to generate a binary string. By selecting discriminative bits in the binary string the final compact local binary descriptor RLBD is achieved. What’s more, Integral image in polar coordinates and polar location mapping are presented to speed up the construction of local binary descriptor.