The Research Of Fast Digital Image Matching

Image matching is a process of matching and overlaying two or more images scene fromdifferent times, different sensors or under different conditions (weather, illumination, cameraposition and angle, etc.) ,it has been widely used in remote sensing data analysiscomputer vision,image processing and other fields. Nowadays, the development trends of image matchingtechniques are faster matching speed, higher matching accuracy and more robust, especially,inthe fields of terrain contour matching, image network information retrieval, moving targettracking and industrial real-time monitoring etc. all need very strong image matching algorithmof real time to deal with it.How to ensure accuracy, improve matching speed have a strongtheoretical meaning and use value. In this thesis, the main research direction is how to achievefast image matching with maintain the accuracy of image matching.Atpresent，the fast image matching method is mainly divided into two aspects which basedon hardware and software. For most of the software-based matching method is based on theimprovement of the algorithm itself, trying to reduce the number of feature for matchingt, or toimprove the matching strategy which reduce the time consumption and accelerate imagematching. The typical hierarchical search technologys are Layered-searching algorithm,Sequential similarity detection algorithm(SSDA), gray combination method, the promoter regionof correlation matching, amplitude sorting algorithm. However, the accelerated degree ofaccelerate image matching is limited which only by improvement of algorithm itself.With therapid development of the GPU and other computer hardware,the processor computing speedcontinues to improvethe, so that using hardware to accelerate image matching is becoming a veryeffective method. Currently informed research is mainly supported by FPGA, multi-core system,the Platform, GPU etc.Based on the above considerations, this thesis summarizes image matching’s developmentand actuality, analyzes the parallel processing and data mapping to apply in GPGPU(GeneralPurpose GPU). The thesis mainly studies the improvement of SIFT algorithm based onSUSAN(SUSAN-SIFT algorithm) and implementation. of SUSAN-SIFT image mathing basedon CUDA. SSFIT algorithm uses SUSAN operator to retain the key points instead of 3Dquadratic fit function and the Hessian matrix，then uses USAN region and extremumproperties of key points to classify key points. Based on image matching algorithmimplemention on the CUDA, refactoring the data structrue of SUSAN-SIFT and implementingthe time bottleneck algorithm SUSAN-SIFT on GPU.The expemental results show that ,firstly, SUSAN-SIFT algorithm can greatly improve thematching speed of the algorithm with the complexity of feature extraction and matching be reduced under the circumstances of high correct image matching rate；secondly,GPU have abetter computing performance than CPU in dealing with lower complexity and large-scale dataand without the data transfer bandwidth, GPU-based image matching method will reach to a veryhigh acceleration. Therefore, SUSAN-SIFT algorithm and GPU-based matching algorithm forfast image matching with a certain significance.