Technologies And Applications Of Visual Saliency Detection For Image Datum

Images are the primary data resource in information society. Voluminous image datum results in the critical challenge for the high efficient information processing intelligently. We notice that the content that a person is interested in is often occupied a small part of an image or a period of video. It is necessary directly to detect the interested areas for high efficient processing results. The processing idea stems from the selective attention mechanism and the perceptual organization principle in the human vision system. Thus, the following items should be dealt with: How to utilize the perception principles of visual saliency? How to describe and distinguish the various saliency events contained in images? How to introduce above psychological theories into the procedure of image analysis effectively? How to extract the salient regions or events rapidly from an image or a video period, which are interested by almost users? This dissertation focuses on these aspects.The first part of this thesis emphasizes on the framework design for visual saliency detection. Firstly, after discussing the relation between visual saliency and image contents based on the theories of cognitive psychology, a new strategy for representing visual saliency is proposed based on content-correlation, by which image salient events can be divided into two classes, low correlative and high. Secondly, a hierarchical framework for describing and understanding image saliency is presented by analyzing the cooperation between attention and organization. Thirdly, an image saliency detection model is developed based on the general attention in order to put selective attention mechanism into the whole procedure of image processing.The second part of this thesis studies on the methods of visual saliency detection for image datum. Firstly, an improved attention driven algorithm for salient region segmentation and feature learning is proposed to obtain salient elements and description for region-based image retrieval. Secondly, the applications on target recognition in remote sensing images are researched. (1) A hierarchical model on man-made object detection is built up and a level set evolution algorithm for man-made region segmentation is developed to focus on salient man-made candidate areas. (2) a man-made object analysis framework based on structure grouping and a method for extracting and grouping line-like structural elements are adopted in order to implement perceptual grouping of man-made configuration. (3) A road detection and extraction method based on classified salient element perceptual grouping is developed. Then a video event detection model based on spatial-temporal attention is presented, and is applied to detect fire events from video images by extracting fire-like salient regions.The final part of this thesis offers a general design method of an experimental system on visual saliency detection in image data, and discusses some potential applications and other relative extend items.The models and algorithms developed in the thesis are applied to various real images and video and the expected results are obtained. It has some feasibility and adaptability.