| 英文摘要 | The technology of specific target detection has great significance in robot navigation, aerial reconnaissance, and so on. Limited by objective conditions, usually, it is very difficult to obtain reference images of the target under different viewpoints. Detection algorithms should exhibit high adaptability to the illumination changes, scale changes and rotation, owing to the considerable discrepancies between the reference images and the real-time images in imaging time, weather conditions, carriers' attitude information and the imaging resolution. In addition, detection algorithms must be highly efficient, because they often need to run in real-time on moving platforms. There are still many problems in existing detection algorithms, for example, low environmental adaptability, lack of stability and inefficiency, which prohibits these algorithms' application in real-world circumstance. In this thesis, we follow a coarse-to-fine strategy and lucubrate on specific target detection based on local pattern analysis. The main contributions are summarized as follows: (1) A fast ROI extracting method based on the target visual saliency model (TVSM) is proposed to reduce the search space. The method makes full use of the information of the target to guild the generation of a saliency map, and extract the ROI through an analysis of the saliency map. Firstly, the local structural patterns are used as features. They can be encoded with a one-dimension number, which enables fast indexing into the weights of patterns. Second, the TVSM is built based on an analysis of the distributions of patterns of the target as well as its environment in the reference image. Based on the acquired TVSM, the saliency map of a real-time image can be obtained by first computing the pattern of each pixel, then indexing into the weights of patterns, and back-projecting the weights. The evaluation of the coarse location of the target is obtained by smoothing the map with a sliding window and taking the position where the maximum value of the smoothed map exists. Experiments demonstrate that the intensity of the saliency map is highly correlated with the target and that the ROI extracted by our method can cover the target in the real-image because of its effectively exploiting the information of the target. Furthermore, the method shows considerable adaptability to illumination changes, blur, JPEG compression, scale changes and rotation, and meets the requirement of high efficiency, ... |
修改评论