Research On Point Pattern Matching Algorithm Based On Invariant Sequence

Automatic Target Recognition (ATR) with computer vision is very important in modernization of national defense technology, production inspection of industry and agriculture and has become a hot research branch. ATR and stable tracking method are two important factors of obtaining advantages in the modern battlefield. In national economy, ATR also has important application, for example:internet image retrieval based content, fingerprint recognition, face recognition, iris recognition, license plate recognition; motion analysis of persons come in or out safety supervision department, component recognition and inspection by robot in industry.When the contour of object can be easy extracted,2D recognition algorithm based on contour is usually used to recognize it. And the matching of two contours is one of the object recognition methods. Point pattern matching (PPM) is to find the best correspondences between two point patterns which are under some geometric transform, so as to recognize objects.In practical applications, object images obtain from different distances, view perspectives and positions. They are always under rotation, translation and scale transformation, even under affine and projective transformations. These make the PPM hard. Many PPM algorithms take long time to find corresponding points, and are limited to similarity transformation, but are not effective under affine and projective transformation.In this paper, a novel 2D recognition algorithm is proposed by research on PPM which is based on invariant sequences and combined with five-point invariance, Fourier transform, gray relational analysis, particle swarm optimization. The feasibility and effectiveness of the proposed algorithm are proved by both theoretical analysis and experiments. The main work in the thesis is described as follows:(1) Image pre-processingImage pre-processing is an important step of object recognition. It will directly affect the correct recognition rate of objects. In this thesis, the image pre-processing includes mainly image segmentation, contour extraction and normalization.(2) Invariant sequence computing Contour is a useful feature in object recognition. Many descriptors, such as, chain-code, Fourier descriptor, the curvature based descriptor. These descriptors are invariant under translation, rotation, and scale, but variant under affine and projective transformation. Five-point invariant descriptor is a contour presentation method and consists of a series of five-point invariants. The descriptor is invariant under similarity and projective transformations, and is also more stable and roust in the presence of spurious points. In this paper, five-point invariant descriptor is used.(3) Point pattern matching algorithmFirstly, extract the contours of model and object images and normalized them; secondly, compute the five-point invariant sequence and do the Fourier transform; thirdly, save all model images' invariant sequences, Fourier spectral characteristics to construct model database; fourthly, Euclidean distance is used to compute the similarity of Fourier amplitude in order to select the candidate models; finally, the starting corresponding point of candidate model contour and object contour is found by GRA, and then the refined corresponding points obtained by PSO and used to compute geometric transformation parameters, so as to found all the corresponding points and recognized the object.(4) Experiments and results analysisWe have selected 31 synthetic objects and 19 real objects as models and obtain images under similarity and projective transformations to validate the effective of proposed algorithm. For similarity transformation, images of 10 patterns with various positions, orientations and scales are photographed for each model. For projective transformation,10 patterns are obtained from different perspectives. The proposed method is also compared with other contour-based algorithms.Theoretical analysis and experimental results show that:(1) five-point invariant descriptor is effective not only under similarity transformation, but also under affine transformation and has better recognition rate than many other invariant descriptors (2) Fourier spectral characteristics can be used to reduce the range of possible models matched, and then reduce matching time. (3) Grey relational analysis isn't strict to data and can be easy used to find the starting point quickly and exactly in point pattern matching. (4)The good characteristic of PSO makes the PPM more exactly and effectively, even under projective transformation. Thereby, the correct recognition rate of objects improved.