Research On Invariant Feature Of Visual Data

With the development of multimedia technology, visual data are exploding. Visual datarecognition, as a key step in the intelligent processing of visual data, largely depends on thequality of the visual data features. In this paper, several key issues in the extraction of visualdata features have been studied, including image restoration, image contour feature extraction,and, the description on static texture as well as dynamic texture. The visual data collectedfrom the real world inevitably involves various types of degradation. Such degradation mightsignificantly affect the stability of the feature extraction process as well as the invariance ofthe extracted features. To remedy this problem, this paper starts with the study on imagerestoration algorithm, which is expected to recover the clear image from the degraded versionand to provide reliable data for image recognition. Contour, as one kind of visual feature thatplays an important role in human vision, has not been fully exploited in image recognition.Compared with contour features, texture features have been extensively studied in the past.However, the computational complexity of the current approaches is still a problem thatshould be treated in practical use. Often presented in the exploding video data, dynamictexture is one type of significant visual feature, resulted from the extension of static texturefrom image space to temporal domain. Compared with static texture, dynamic texture canprovide additional cues for recognition. Nevertheless, the statistical self-similarity of dynamictexture along temporal axis has not been fully exploited. Motivated by the problems above,this paper studies the contour feature, the static texture feature and the dynamic texture featurerespectively, with several effective and robust solutions proposed. The contributions of thispaper are concluded as follow:1. A scheme for constructing a non-local wavelet frame as well as a wavelet tight frameis developed, based on which the sparsity regularization can simultaneously exploit both thesparse prior of local variations of image intensity and the non-local self-recursive prior ofimage structures over the image. Built upon the proposed construction scheme, a powerfulregularization-based method is developed for solving general image restoration problems. Theproposed method can perform well on both the cartoon-type region and the texture region. 2. By introducing the concept of physical torque into image space, a significance metricfor contour patch is proposed. Based on the torque-based metric, a powerful contour-relatedfeature is proposed with a novel contour patch detector and a robust contour patch descriptor.Integral image trick is used to accelerate the calculation process of the proposed feature. Theproposed contour-related feature can perform well in classifying complex objects, and is ableto provide complementary information to texture feature.3. Combining local binary patterns coding with global fractal analysis, a powerful texturedescriptor is proposed and applied to texture classification. Compared with current localbinary pattern based approaches, the proposed approach can perform well in recognizingrandom textures from the real world. Compared to the state-of-the-art fractal-based methods,the proposed approach shows its advantages in speed, feature compactness and classificationaccuracy.4. A spatio-temporal multi-fractal analysis scheme is introduced for extracting dynamictexture features, which consists of two sub-schemes: one is the volumetric analysis for globalcharacterization and the other is the multi-slice analysis for encoding the local fractalbehaviors of dynamic texture along different axes. Combined with four spatio-temporalmulti-fractal measures, the proposed approach can fully capture the self-similar structuresexisting in dynamic texture from different perspectives.