Visual Attention Modeling For Multiple-instance Images And Visual Saliency Applications

As the rapid development of image acquisition equipment and multimedia technology,the scale of image data increases every day.Visual analysis faces the huge image data.Although the computer calculation ability is very strong,the computer image understanding ability is far less than human visual ability,because of the semantic gap between image features and high-level semantics and the subjectivity of user.The computer doses not extract,analyze and understand the image information according to human visual attention during image analysis.Human visual perception system can select some important contents from huge amounts of information.The selectivity and initiative of the physiological and psychological activity is known as the visual attention mechanism.Human observation image is a knowledge reasoning process of visual perception.Usually,an image contains a variety of target information,while human visual perception system can learn the correlation between image contents from multi-instance images to obtain the interested target.Obviously,the multi-instance images visual attention is introduced in the fields of image analysis and processing,which can select interested area in images quickly and accurately.It has important research significance for the visual search task efficiently.The existing visual attention models are mostly for single image,do not automatically learn the correlation between image contents from multi-instance images,and can not correctly simulate the process of human subjective visual attention.Therefore,aiming at this problem,the thesis studies the visual attention model for multi-instance images.We propose a new visual attention model based on multi-instance learning for multi-instance images.Inspired by human visual sense,based on recent research about visual attention model,we focus on the research about visual feature saliency measure,feature saliency fusion for multi-instance images,saliency object extraction and visual attention model for multi-instance images.At the same time,we research on visual saliency application in the field of image processing and analysis.We propose a single image retrieval method based on the color saliency distribution histogram of bit plane and a multi-instance image retrieval method based on salient object respectively.Introducing visual saliency to image fusion,we propose a novel algorithm based on visual salience in wavelet domain.Our algorithm can achieve more clear minutiae and improve fusion performance of human vision.It is also benefit to object detection and recognition.The contributions of the dissertation are listed as follows:(1)To solve the problems of incomplete feature types and poor anti-noise of visual attention model,the process of human visual perception is analyzed according to different image features.Image features are combined with human visual perception,and the features saliency measurement method based on visual perception is proposed after consideration of psychological factors.This method is designed and extracted to simulate the visual stimulation with the principal bit plane intensity feature,HSI space color feature,Gabor direction feature and multi-directional texture feature.Compared with the existing visual attention model,it is more comprehensive and detailed in the description of image visual features,more in line with human visual characteristics,and improves the anti-noise performance of visual features.We measure different feature saliency from self saliency,global contrast,local contrast and sparsity saliency.The local contrast saliency measurement based on wavelet transform is designed to make feature saliency more in line with human visual perception process.(2)A multi-instance image visual attention model based on multi-instance learning is proposed.By analyzing the correlation between images and visual features,the model designs feature selection based on multi-instance learning and feature saliency fusion algorithm based on attention to form comprehensive visual saliency map and extract salient objects from the map.Contribution of a feature saliency map includes two parts.The one is the relation between feature and images.The other is the self prominence of feature saliency map.Then according two parts to fuse calculation,we weight different features' contribution to the final image saliency.At the same time,a fuzzy region growth algorithm based on global optimization is proposed to extract salient objects and avoid excessive segmentation.Compared with the existing visual attention model,the proposed model fully considers the correlation of multi-instance images and improves the integrity of salient objects.The proposed model better reflects the user's interest object,and subjective visual attention is closer to the human visual process.(3)Introducing visual saliency to content-based image retrieval,we propose a single image retrieval method based on the color saliency distribution histogram of bit plane and a multi-instance image retrieval method based on salient object respectively.These methods removed a large number of backgrounds of image,improved precision ratio of image retrieval.The method based on salient object of multi-instance images can better reflect the user interest object.(4)Introducing visual saliency to image fusion,we propose a novel algorithm based on visual salience in wavelet domain.Visual saliency model has been introduced into calculating the intensity,direction and texture global integrated saliency of each pixel at each level to adaptively acquire the weighted fusing coefficients based on multi-scale wavelet decomposition.Experiments show that our algorithm can achieve more clear minutiae and improve fusion performance of human vision and objective evaluation,comparing with regional feature-based wavelet fusion algorithm.It is also benefit to detection and recognition.