Top-Down Attention Motivated Research On Perception Model

During the last few years, extraordinary progress has been made in understanding the basic principle of how information is processed by visual cortex. This brings more and more concerning to bottom-up attention at home and abroad, and models based on it are built successfully. But difficulties are met during the study of top down attention guidance. Most of current top-down attention motivated perception models were derived chiefly from former bottom-up attention based models. These models lie on data that comes from bottom-up attention and lacks of effective definition of high level information. Hence studying top-down attention motivated perception theory and building the corresponding model have become a problem that needs urgent solution.The perception of object and spatial position is a main finding in the study of visual perception system inspired information processing theory. Its virtual study is about the function of two visual subsystems—"what" and "where" pathways. Accordingly, inspired by the theory of two visual pathways, a top-down attention motivated visual perception model is built in this paper. The integration of bottom-up and top-down process will affect our attention in normal human visual. So the top-down attention perception which we emphasize is actually a process of interaction of top-down and bottom-up.The paper classifies perception of spatial position related to "where" pathway into two levels: spatial context of object and its specific position in visual space. We use the contextual information as the first level "where" information transmitted in "where" pathway and use the specific position of object in visual space as the second level "where" information. Besides, we use the perceptual information about object as "what" information transmitted in "what" pathway. Integrated with attention mechanism, the first level "where" information can motivate top-down attention and provide guidance for object perception. Bottom-up attention is motivated by the second level "where" information and "what" information.The main innovative points of dissertation are as follows:First, inspired by research of visual attention in psychology, a novel algorithm for extracting bottom-up attention information (Integration of local complexity and early visual features, LOCEV) is proposed in the paper. Bottom-up attention information is composed by saliency of certain regions correspond to each point in image, and scale of the regions varies with complexity of local features adaptively. New saliency metric is defined as a product of three terms: local complexity, statistical dissimilarity and early visual features. Salient regions are salient both in feature space and over scale. Saliency of certain regions correspond to all points in image is defined as "what"information. The position and scale information is defined as the second level "where" information. The extracted bottom-up attention information is invariant to image scale, rotation and translation, and is shown to be robust to noise.Second, a novel algorithm is proposed for extracting context-centered first level "where" information (CONCEN). Three procedures are carried out in CONCEN algorithm: 1. extracting high dimensional code of the first level "where" information; 2. sub-sampling the high dimensional code of the first level "where" information; 3. computing statistical feature of sub-sampled high dimensional code of the first level "where" information. The final coefficient of statistical feature is defined as the first level "where" information. Integrated with attention mechanism, the first level "where" information can motivate top-down attention and process contextual information over a large area. The information of entire scene is coded in the first level "where" information. It can provide reliable prior knowledge for bottom-up attention.Third, inspired by the theory of two visual pathways, a novel visual perception model is proposed based on "what" and "where" information (WHAT-WHERE). Context-centered first level "where" information is used to control top-down attention, and guide bottom-up attention which is driven by "what" information and second level "where" information. The procedure of top-down attention can be divided into two stages: pre-attention and focus attention. In the stage of pre-attention, first level "where" information can be used to provide prior knowledge of presence or absence of objects which decides whether search operation is followed. By integrating the result of focus attention with "what" and second level "where" information, attention is guided to the region that is most likely to contain the object and series of salient regions for samples are detected. An attention model is developed based on these. Experiments results with natural images demonstrate its effectivenessFirst level "where" information can provide strong priors to tell whose presence is most possible and where it will appear. Besides, it can help to disambiguate the identity of the object while lacking of the local features. The entire detecting process may be stopped only after pre-attention. And the result of focus attention can be integrated with bottom-up attention to provide efficient mechanism to focus attention on salient regions.