Research Of Salient Regions Detection Based On Bottom-Up Visual Attention Model

With multimedia information technology research, the research of how to understand image is taken more and more attention. Thus, human visual attention model is gradually becoming the hot spot in the field of multimedia processing, including the application on image content analysis, image retrieval, image quality evaluation, compression, scene monitoring, image/video adaptive and so on. Therefore, how to build visual attention model which is more in line with the human eye of characteristics of attention will become the focus of research at home and abroad.At first, this thesis introduces the basic structure of knowledge of the visual attention model, including the human visual system, visual attention mechanism and visual attention models based on two different visual mechanism; And then, the thesis mainly introduces a visual attention model based on multi-scale local contrast of low-level features and based on this model, a visual attention model with cross-layer saliency fusion is proposed; The thesis also focus on the application of visual attention model which automatically extract salient region of image and video, including the application on image retargeting and the preliminary application on video monitoring; At last, this thesis introduces the improvements on the existing classic Stentiford visual attention model and potential applications on image copy detection.The main contributions of this thesis can be summarized as follows:Firstly, this thesis proposes a visual attention model based on multi-scale local contrast of low-level features. The model is proposed based on the construction theory of Itti’s model and has better performances than Ittis’. Instead of the center-surrounding difference computation, we adopt multi-scale local contrast of low-level features to extract image salient region. Experiment results show that the proposed model has better performance than Ittis’Secondly, this thesis proposes a visual attention model with cross-layer saliency fusion. In the process of constructing model, not only the local feature contrast, but the global feature contrast is considered taken part in forming the salient map. And different from the feature fusion method that simply linear addition of global and local features, in this method, based on the global priority principle of perception, the global and local features are firstly combined to form a weighted model, and then in turn to weight the global model to get the final salient region. Experimental results demonstrate that the proposed model performs competitively with four existing models on detecting out accurate salient regions and enhancing the contrast between salient and non-salient regions.Thirdly, this thesis introduces the application on image retargeting based on the visual attention model with cross-layer saliency fusion. Based on the proposed visual attention model with cross-layer saliency fusion, this thesis puts forward a novel image retargeting method and dynamic browsing strategy;Fourthly, this thesis introduces the preliminary application on video analysis based on the visual attention model automatically extracting salient regions. Based on the spatial saliency map, we introduce temporal factor (movement characteristics), and through effective integration to get the spatial-temporal visual attention model. The thesis also introduces its application on video scene monitoring and gets satisfactory results;Fifthly, this thesis proposes the improvements on the existing classic Stentiford visual attention model. At the same time, this thesis uses the improved attention model to all sorts of copy processing images to extract salient regions, and the extraction of salient regions has good stability, which demonstrates it can provide valuable reference features to image copy detection.