Infrared Target Recognition Method Based On Local Similarity Structure Statistical Matching Model

With the high-speed development of infrared technology and computer vision technology, infrared target recognition in image has widely been applied in various fields. At present, the methods of target recognition based on model-matching and learning classification can obtain good recognition results, but they are highly dependent on template and training sample set, and weak ability of generalization for general infrared targets in the complex infrared scenarios. To solve this problem, a training-free target recognition method is researched, by introducing the idea of local matching, and using simple template set, a robust infrared target recognition method based on local similar structure statistical matching (LSSSM) model is proposed in this paper.In the research of infrared target recognition method based on the training,a feature extraction method of locally adaptive regression kernels (LARK) is studied in depth, and it is applied to extract the local structure of infrared image in this paper. A simple template set containing few images of an interested infrared target is built, and a dimension reduction method is proposed in this paper by eliminating similar structure of the template set to obtain independent non-similar structure features set of the interested target. In order to achieve the robust recognition for the infrared target with complex structure, a local matching model is presented in this paper, by counting the number of similar structures between one area of test image and simple template set to generate the similarity image of the test image and the interested target. In this paper, the method of non-maxima suppression is researched, and this method is used to extract target information in the similarity images, then, marks the target in the test image.The method of LSSSM model in this paper, compared with the method of template matching and training, it doesn’t need comprehensive template and complicated training process. The experimental results show that the present method has robust recognition ability for typical infrared targets such as infrared human body and infrared cars in complex night vision scenarios.