Research On Extracting Texture And Blur Invariant Features

Features extraction and features representation play a key role for image retrieval、 image classification and object detection which are major content in computer vision. The effective method of features extraction can improve the performance of retrieval、 classification and detection, furthermore, the features which are chosen are decided by the exact application. In this dissertation we focus on the features extracted from image and construct the gradient orientation and modulus matrix (GOMM), the histogram of self-correlation (HOS) and the local phase pattern (LPP), which are attracted from the gradient domain of the image, the time domain and the frequency domain respective. And they are used for texture classification、human detection and blurred image classification respectively.The GOMM contains the spatial correlation information between pixels in specific direction, but the features for the texture are not accurate and comprehensive enough. The gradient of image reflects not only the variation intensity of pixel, but also the variation direction of pixel. So the gradient is the new method for the features extraction. In chapter2, the gradient is firstly calculated, and then gradient orientation and magnitude are further quantized into several dominant orientations and main levels respective. The GOMMS are constructed from main levels of gradient magnitude; the number of GOMMS is decided by dominant gradient orientation. At last some statistic features are extracted from GOMMS. In the following experiment, we evaluate the GOMM on texture database, which has multi-scale, multi-viewpoint, different light condition and different materials. The GOMM shows the best performance among all image descriptors used in the experiments.HOG features include gradient orientation and gradient magnitude information, these features which are trained by machine learning algorithm (such as support vector machine) reveal the object’s contour. Inspired by above method, we design a new features extraction method, in which the structure features are derived from the self-correction surface. In chapter3, the correlation surfaces for each pixel in the local patch is calculated from the correlation value of a block, then the correlation surfaces are transformed into log-polar coordinates. Thirdly log-polar coordinates is divided into several areas, in which the log-amplitude and angle are equally partitioned. The maximum value is reserved for each area. At last the number from each area is concatenated into the histogram of self-correlation (HOS) for the local patch. In order to get the highest accuracy of classification, the parameters of HOS are changed. In the person/non-person classification experiments, HOS outperforms HOG.In practice, the images are often blurred, the blur invariant features are proposed in chapter4. It can be proven that the tangent of the phase is invariant when the image is transformed into frequent domain. The local phase pattern (LPP) is derived from the local tangent of phase for each pixel. At last the histogram of LPP is constructed for the patch in the time image. In the following experiments of blurred texture and face recognizing, LPP achieves promising object classification rate, even when the amount of blur significantly increases.