Target Recognition Method Based On Structure Of Local Feature

The problem of target recognition lies at the heart of computer vision research. The key technique lies in image recognition is feature. More and more researchers focus their researches on local feature due to its good stability and discriminability. This paper researches target recognition method based on structure of local feature, which introduces the structure of image local region with the appearance information. It is able to dig inherence relationship of target and improve the modeling precision. The main work includes:1. Reviewing a variety of feature detectors and descriptors. Based on the traits of key-point regions detected by feature detector, three complementary types of detector were chosen: Kadir&Brady, DOG and edge detector. We give details about each dataset used in the paper, commenting on its difficulty in application.2. Target recognition method based on structure of local feature is analyzed in detail. To resolve the problems of simple shape model, we research full model and simplified star model in which appearance, shape and scale are all modeled by Gaussian probability density functions. They describe relationship of target more clearly than simple shape model which combine appearance with shape learnt in two separate stages. Besides, based on Probabilistic Latent Semantic Analysis, feature region representing location information is added, which give a try in depicting objects over latent indefinite space relation. This approach models histograms of regions, so it's a dense model and theoretically copes well with occlusion.3. Analyze the two approaches: shape model approach and adding location into pLSA approach. The matters in modeling and effect factors are detailed including feature detector, number of parts and detections, contribution of different model terms and result analysis of more challenge datasets.4. Address the proposed algorithms of target recognition to aircraft target in remote sensing image. The stability and discriminability of local feature detected from remote sensing image are validated. Then, local feature based recognition algorithm for high resolution aircraft target is proposed. As a part of it, the curtly locating method use grid. It provides a good performance in throwing off false alarms and identifying aircraft type.