The Scale Optimal Selection For Image Global Representation

We live in a world of targets with different scales. These difference indimensions is determined by the properties by the object itself, hierarchicalrelationships between entities, as well as the perspective projection of the humaneye receives information, etc. In the field of computer vision, the informationreceived by our human eyes is distributed in a certain range of scales. When we arereceiving these information, and understanding these information as meaningfulcontents, we will research them on different scales, and correspond its specificinformation under different scale with different scales of high-level abstractions.And then those computer vision tasks can be processed, such as image recognition,scene classification, and many other computer vision tasks.The analysis of image scales has become an inevitable issue in the researchwork. The multi-scale representation in image scale analysis and scale selectionwork has been analyzed systematically, and for the purpose of global image scaleselection, from the aspects of the continuity of the image, the image characteristicsof repeatability, analysis of the scale of the scene images within the global scope, aswell as the saliency theory, it is believed that the optimized scale in the global scopeexists under some kind of constraint. At the same time the global optimal scaletheory was put forward, and was augmented and elaborated.Scale selection and feature selection is inseparable, and scale selection workmust be merged into feature selection to have the meaning of visual understanding.However, scale selection in feature extraction is mainly used in multi-scale interestpoint detection, however, the scale selection in dense sampling has been ignoredwhich has been another increasingly important strategy in many computer visiontasks. Dense sampling has a large amount of information at the same time, and theimage sampling under a constant scale can be seen as the scale of the image in theglobal scope.The images consisted of different objects under different scales can be seen asthe mixed data from different subspaces, the subspace optimization problem can betarget into a subspace segmentation problem. Low rank expression was used, whichhad very strong robustness, was not affected by the fact that it could bring datadamage when only one size was adopted in every feature, and the local consistencycan be enhanced effectively. And augmented Lagrangian multiplier method is usedto solve the low rank problem, solving the optimization problem in a fast way.The proposed global image scale selection method based on low rankrepresentation got outstanding results in classification applications, such as scene images and clothes images tasks. The image scale theory in global scope was firstlyproposed; the absence of scale selection work for dense sampling feature extractionwas made up, and the scale selection method was proposed and formulized; theglobal scale optimized problem was turned into subspace segmentation problem; theperformance of feature expression, especially for image classification was enhanced;the existence and rationality of this theory was discussed in theory and experiments.