Research For The Selective Attention Mechanism In The Image Information Processing

Computer image information processing has been a focus of attention in modern information era, because traditional manual image information processing is confronted with many difficults while the amount of images that are the main carriers of information at current becomes more and more prodigious. We notice that the interested content is often a small part of an image in the task of image analysis and understanding. So, it is urgent to import and research the selective attention mechanism in human visual information processing in order to increase the effectiveness of computer image information processing.These are three primary aspects in the research of selective attention mechanism for image information processing: How to construct a practical framework for the selective attention mechanism? How to automatically identify the focus of attention (FOA) that generally is the interesting area in an image? How to efficiently perform the current task of image analysis and understanding under the direction of the selected FOA? This dissertation studies these aspects in detail, and some valuable results are achieved.The first part of work in this thesis focuses on the framework design. An approach of image information processing including the selective attention mechanism is framed based on some psychological theories about the human visual system. The procedure of image information processing is functionally divided into three stages: the pre-attention stage at which the image information is collected and coded, the attention stage at which the FOA is detected, and the post-attention stage at which the FOA is explained.The second part of work in this thesis focuses on the FOA selection, and describes the detailed implementation of the pre-attention stage and the attention stage. At the pre-attention stage, an approach of simple image feature extraction is presented, in which the single original image is decomposed into the multiple separate maps for separate properties in parallel. At the attention stage, bottom-up and top-down FOA detection are studied respectively: An approach of bottom-up FOA detection that is independent of the task is proposed, in which the competitions of scale, feature, location and size are implemented hierarchically to define the fundamental attributes of the FOA, and a series of FOA that are novel in the image are successively detected in this way. Also, an approach of top-down FOA detection that is dependent on the task is proposed, in which a FOA model that regulates all kinds of competitions in FOA detection is structured according to the task, and then a series of FOA that are relative to the current task are successively detected.The third part of work in this thesis focuses on the FOA application, and describes the detailed implementation of the post-attention stage. Firstly, the selective attention mechanism is applied to the image segmentation, and a novel segmentation algorithm is brought up, which combines the approach of FOA selection and the traditional method of region growing. Secondly, the selective attention mechanism is applied to the object recognition, and a novel recognitionalgorithm is brought up, which combines the approach of FOA selection and the traditional symbol-based method. In the two algorithms, the computing resources are assigned under the direction of the selective attention mechanism, which reduces the amount of processing and increases the efficient of image analysis.The proposed approaches and algorithms are applied to the various real images respectively, and the.experimental results are prospective.