The Study Of Image Objects Description And Its Completeness Based On Local Features

With the application and development of Internet technology, peopleare exposed to a wide variety of images every day. Facing this mass ofinformation, just dealing with them artificially is obviously unrealistic.Along with the increasing power of computer processing and thedevelopment of computer science, local invariant feature in patternrecognition gets more and more attention owing to its excellent stability,repeatability, and matching ability. However, in practice, the number anddistribution of the local invariant features are still uncertain. There exists noreasonable scheme till now. Based on the knowledge of the local invariantfeatures, this paper tries to explore the completeness of these features whichmay provide a more reasonable solution.This article begins with the scale-space theory and affine invarianttheory, then comes up with an overview of the technical processes of localinvariant features, and then detailed studies several common methods ofdetecting and discripting local invariant features, such as MSER, SIFT,ASIFT. On the basis of the above study, this paper focuses on the completedescription of local invariant features. Through the analysis of a simplebinary image, a general approach of getting the complete description isobtained. By extracting the life of each local neighborhood named skeleton in the gray level difference space, the importance level of each localneighborhood is distinguished, which provides the complete description ofthe entire image. To be more directly and clearly, the complete descriptionis showed in the three-dimensional space.Experiments based on the complete description proposed in this paperclassify two data sets using BOF model, aiming at verifying the correctnessof the theory. As a comparison, eight other experiments using the samemodel with eight different feature extracting method are performed, that is,uniform sampling, SIFT, MSER, Harris Affine, Hessian Affine, IBR, EBRand another complete discription simulating the process of imagecompression. Experiments achieve satisfactory results in both of the twodata sets.