Research Of Image Group Scanpath Generation And Prediction In Natural Scene

Thanks to the regulation and control of Visual Attention Mechanism(VAM),the human eyes,when viewing a visual scene,tend to quickly focus on Region of Interest(ROI)in the whole scene.If the human Visual Attention Mechanism can be effectively modeled,and the corresponding model can be successfully applied to the field of modern computer vision,the computer image processing algorithm will understand the image scene content to a human-like level,and further deal with a variety of complex visual tasks efficiently.Group scanpath generation and prediction of image is a research topic,which born for modeling and simulating the human Visual Attention Mechanism,and especially its dynamic transfer process.After a long period of development,some related breakthroughs and progress have been made in the fields of biological neuroscience and cognitive psychology.With the rise of big data mining and artificial intelligence,researchers have begun to use machine learning technology to carry out some new explorations.Group scanpath generation research usually uses data mining technology to analyze the eye tracking data of group observers,and determines the common visual elements and visual attention shifts patterns.The scanpath prediction research usually refers to the physiological principles of the visual system in the fields of biological neuroscience and cognitive psychology,and further builds model-driven or data-driven mathematical models to simulate the process of human visual attention shifts.At present,the relevant research of this topic has a large space for development,needs and deserves further exploration.Inspired by the related research in the field of neuroscience,cognitive psychology and computer vision,this thesis uses the eye tracking data of group observers to mine the human Visual Attention Mechanism in natural scene,and explores the topic of group scanpath generation and prediction.The main work contents are summarized as follows:(1)The thesis study the topic of group scanpath generation in natural scene,and proposes a group scanpath generation method based on region of interest clustering and shifting.At first,the fixation points are clustered to determine the regions of interest;then the selection of representational regions of interest,the generation of gaze behavior related sub-regions of interest,and the gaze priority related transfer mode of regions of interest are further studied;combining the prior information of scanpath length with the above process,a group scanpath is finally generated.(2)The thesis study the topic of scanpath prediction in natural scenes,and proposes an prediction model of visual attention shifts across instance objects based on nonlinear regression.The model uses the state-of-the-art segmentation algorithm to extract all instance objects in the scene;then it combines eye tracking data of group observers to extract the visual attention priori features of the instance objects;using nonlinear regression neural network to fit the correlation between the priori features and fixation priority,it predicts a visual scanpath who's fixation points are represented by instance objects.