Research On Chip Design Of SIFT Algorithm With Its Image Feature Processing

SIFT (Scale Invariant Feature Transform) algorithm is one of image feature processing algorithms to complete feature extraction and description. It is based on the observation of scale space and diverse invariances of images in this space. By simulating the changes of visual angles when people observe objects from different distances, SIFT algorithm can achieve more precise descriptions of image features when compared with traditional ways. It also has many characteristics, including scale invariance, rotational invariance, illumination invariance, affine invariance and noise immunity to a certain extent. With these advantanges, SIFT algorithm has improved the accuracy of relevant filed, such as image matching and localization, image registration, pattern recognition and tracking, and generated many other similar algorithms with more outstanding performaces. Therefore, it is significant to put the SIFT algorithm into research and practical use.With the development of VLSI and FPGA, the calculation of image processing could be realized on hardware platform with its tremendous register resources and faster speed than software system. By transplanting the SIFT algoirhtm to hardware platform, we could extend the applications of this algorithm, enhance the processing abilities of relevant integrated systems, and promote the practice and devemploment of correlated mathematical theory.This paper mainly discuss the design of hardware system of SIFT algorithm. Based on the mathematical theory proposed by Professor Lowe, standard program of software system in OpenCV, and achievements of scholars at home and abroad, it chooces the "feature points extraction" part of SIFT algorithm as the research object and propose its hardware system architecture. The main contents of this paper are:Firstly, this paper summarizes the whole process of SIFT algorithm. It analizes and emphasizes those parts of the alogorithm, which have low attention or inaccurate interpretation, including "scale space and its principle of continuity" and "Gaussian function and its parameters' setting";Secondly, this paper summarizes the design of calculation of SIFT algorithm in OpenCV, along with the adjustment and restriction of algorithm in this design. It also adjusts and optimizes the algorithm furtherly on each step and propose the suitable process of calculation, considering the requirements and constraints of hardware environment. The main adjustments include:fixing the size of input image, limiting the precision of data, changing the generation way of Gaussian pyramid, fixing the size of all Gaussian templates, removing the process of "accurate positioning of extreme points" and "eliminating unstable extreme points", changing the formula of gradient calculation;Thirdly, this paper presents the importance of reasonable hardware structure for SIFT algorithm, and propose using "pipeline structure" as the final decision. By using "pipeline structrue", a system could utilize the register resources to the maximum extent and reduce the processing time of the system. This paper also proposes the idea of "modular design" innovatively, considering the similarity between each step of algorithm. This idea aims to split the "pipeline structure" into basic modules with different functions. These modules are designed as templates and could be easily instantiated. Both of "pipeline structure" and "modular design" emphasize the precise control of the time. This paper proposes the final design of hardware system of SIFT algoirhtm with these ideas;Finally, this paper designs integrated tests to verify the performance of proposed system, through combining with the software platform. The results of tests demonstrate that, the proposed system has the characterstics of real-time processing, low resource consumption, high matching accuracy and satisfying noise immunity, which prove that this system meet the practical requirements. In addition, this paper also analyzes the influence of the calculation adjustment. The results of tests demonstrate that, although the setting of size of Gaussian template and precision of data will change the distribution and number of feature points, the matching accuracy will not be disturbed usually, which proves that SIFT algorithm has extremely strong ability of robustness and tolerance of faults.