Multi-angle Based Stereo Information Optimized Acquisition And Object Recognition

With the development of remote sensing earth-observation technology, not onlytraditional types of object information could be acquired such as radar signals,optical image and etc, but also makes it possible to obtain stereo information whichcontains optical characteristics and three-dimensional structural properties of object.Remote sensing stereo information subvert the traditional concept of remote sensinginformation, while retaining optical properties comprehensive reflected by ofsurface material and environmental, more emphasis on sense of presence in athree-dimensional of the object,which more truely reflects the nature property of theobject. It changes from just in sight before to now at your fingertips. Fulllyacquisition and analysis of remote sensing stereo information of the object, canfundamentally solve many application problems which traditional remote sensinginformation could not do owing to lack of describing ability. Digging the potentialapplications of stereo information will greatly enhance effectiveness ofimplementation for many fields, including urban planning, geographic exploration,earthquake disaster reduction, as well as military reconnaissance. They all havemajor significance for the development of national economy and national defense.Thus, it is a urgent mission to establish a complete theoretical system for remotesensing stereo information processing. Optimization of stereo informationacquisition, stereo feature extraction and object recognition are three keytechnologies of the theoretical system, which corresponding to processes ofacquisition, analysis and application of stereo information, respectively. Differentfrom remote sensing classification, object recognition more emphasis on the finedistinction between the similar objects. Therefore, feature with higher describeability and better performance of classifier are needed. Techniques of optimizationacquisition and feature extraction are for increasing the ability to describe the object,and optimization of classifier is to improve its performance for object recognition.Obviously, it will make an indelible contribution for improving the theoreticalsystem of stereo remote sensing information processing, if above problems aresolved.First of all, remote sensing stereo information of object could be divided intotwo parts which are three-dimensional structure and optical information of thesurface. For full acquisition of surface optical information part (the word ‘full’means complete and high quality), a multi-angle based optimized stereo informationacquisition method is proposed. The method solves the planning optimizing problemfor multi-angle observing, which including optimizing work for number of observing points and observing angles of each observation point. In this way,efficient multi-angle based information acquiring program could be obtained, whichcould decrease the loss of information caused by bad observation angles. Owing tothe objective function for this optimizating problem is difficult to build, adequacyevaluation model for multi-angle based stereo information acquisition (AE-MSIA) isproposed. The model can effectively describe adequacy of stereo informationacquired by different types of multi-angle programs. Furthermore, to solve thismulti-variable optimizing problem, artificial bee colony algorithm is utilized.Experiments show that for different types of objects, efficient multi-angle basedobserving programs could be obtained by method proposed in this paper, and mostefficient observing points and observing angles of each point are got. Proven,adequacy of stereo information could be significantly improved under optimiziedobserving program.Further, basing on full acquisition of stereo information of object, to improverecognition accuracy, feature extraction technique is to excavate key attributedescription which can effectively characterize the object. For poor performance oftraditional features for target description, stereo feature including opticalcharacteristics and three-dimensional structural features of object is proposed. Andfocusing on the traditional remote sensing three-dimensional feature extractiontechnology can not adequately describe three-dimensional structure of object, basedon the theory of spherical harmonics,‘spherical harmonic descriptors’ is proposedfor remote sensing three-dimensional information based feature extraction.Moreover,‘3D-Zernike descriptors’ based feature extraction method is introduced to solve theapplication proplems of spherical harmonic descriptors, such as radius ofdecomposition sphere is difficult to choose, large volumes of data, noise-sensitiveand etc. Spherical harmonic descriptors and3D-Zernike descriptors couldadequately describe three-dimensional structure of object. They are conducive todescribe the differences between objects with different three-dimensional structures,and they have rotation invariance in three-dimensional space, which is conducive todescribe consistency between same objects in different coordinate systems or withdifferent towards. The superiority the three-dimensional feature extractiontechniques is proved by comparative analysis. In addition, the importantcomplementary role of optical features is verified, which prove that the stereofeatures can effectively increase the degree of difference between different objectsand the degree of similarity between same objects, which provides more reliableclues for high similarity objects recognition.Finally, remote sensing target recognition technology is to accurately judgecategory of objects based on the extracted features. To solve the poor performanceproblem of traditional features based high similarity recognition, on the one hand object recognition method based on high-performance stereo features to improve therecognition rate of high similarity recognition; On the other hand, a classifieroptimization method is proposed based on artificial bee colony algorithm,whichfocused on when doing recognition with poor stereo features, conventional classifierparameter settings could hardly guarantee recognition accuracy. In this method, byparameter-vector optimization of support vector machine, position of thegeneralized optimal hyper-plane is adjusted, and performance of the classifier couldbe improved. Furthermore, when facing multi-category recognition problem,traditional optimizing algorithm based parameter-optimization of classifier lead tolow efficiency, dynamic artificial bee colony algorithm is proposed for supportvector machine parameter-optimization. This method can effectively improveproblems of slow convergence and solution quality, which are caused by largerdimension of optimized parameter-vector. Experiments show that the proposedmethod could improve performance of classifier. Especially for multi-categoryrecognition problems, compared with traditional heuristic algorithms, the proposedmethod significantly improves the efficiency for parameter-optimization, andeffectively improve performance of classifier, and provides a reliable guarantee foraccuracy of object recognition.It is worth noting that this study exceed the idea of scene based work in thetraditional remote sensing information processing, and completely target individuals.Not only it conforms refinement and stereo trends of remote sensing information,moreover, it provide a basis for future development of theoretical system for remotesensing stereo information processing.