Computer Model Research Of Visual Attention Based On Cooperative Work Between Spatial And Object Attention

Visual attention is the psychological phenomenon which carries attention selection with visual information. It is closely linked with our life and provides guidance to us for selecting a few salient visual regions quickly to pay more attention to useful information. Computational model of visual attention can assign the limited processing resource in system priority to a few salient visual regions. The research of visual attention computational model can push the development of many computer fields such as image understanding. For image and video compression, we can get high compress degree but good observation effect image and video with the method to compress those important regions which on the visual attention focus with low compression ratio on one hand, and to compress other regions with high compression ratio on the other hand. For target searching, increase the typical characteristic's weight of searched target can find the region which the searched target most likely in quickly before target recognition, this can improve the efficiency of target searching. In a word, the research of visual attention computational model has important significance to many fields such as automatic target detection, image and video compression, robot scene classification and medical image processing.The object attention and spatial attention algorithms are studied separately in dissertation based on the research results in visual attention of psychology, physiology and neuroscience. Then a computer model of visual attention based on cooperative work between spatial and object attention has been build finally. The main contents in this dissertation can be summarized as follows: (1) Analysis and summary of the research status of visual attention computational model at home and abroad was made, the structure mechanisms of some classical visual attention computational models at present were discussed from aspects of spatial attention models and object attention models respectively. According to the latest research results of visual attention in psychology and cognitive neurology, the thinking that spatial attention and object attention are on the same grade and work with each other in visual information processing is accepted. And based on this mind, we proposed to build the visual attention computer model which based on the cooperative work between spatial and object attention.(2) For visual attention depends on visual information processing system on physiological structure, the brain mechanism of visual information processing system was researched. Based on the analysis of each layer's physiological structure and function on primary visual pathway, with many visual information processing characteristics such as the receptive field characteristic of biological cells and parallel treatment mechanism of different attribute visual information, the idea that the low layer visual information should be both serial and parallel extracted with the spatial and object model built was explained.(3) On the object attention algorithm aspect, according to the deficiency of current object marshalling algorithms, a new object marshalling algorithm based on edge detection with maximum gradient and c-means clustering algorithm was proposed by combining the merits of edge detection algorithm and color clustering algorithm in object edge detection. The proposed object marshalling algorithm was used to realize auto and quick marshalling of objects in visual scene.Based on the proposed object marshalling algorithm, a computer model of directional visual attention with three-layer structure was build to simulate object attention. It provided object attention foundation for building spatial and object attention cooperative work model. The proposed model transformed visual image into retina image firstly, then obtained all objects'fundamental features in retina image with object marshalling algorithm based on edge detection with maximum gradient and c-means clustering algorithm, the model controlled the shift of visual focus according to the characteristics of directional object in knowledge database and the saliency of all objects in the scene lastly. Relevant experiment confirmed the validity of the model and it's rationality in cognition.(4) On the spatial attention algorithm aspect, many classical algorithms at present on the three main stages which are feature vector extraction, activity image computation and combination in the processing of spatial attention saliency image computation were discussed. According to their emphasis points and deficiency, we improved these three main stages'algorithms respectively.On the low layer visual feature vector extraction method aspect, inspired by the research in physiology, a retinal transformation method based on contrast sensitivity was proposed. The original image is weighted with a contrast sensitivity formula which is a function related to retinal eccentricity before intensity, color and orientation feature vector extraction to obtain retinal transformation image.On the activity image computation and combination aspect, an activity image computation method based on Markov chain was proposed by improving the method which built Markov chain on complete graph proposed by Harel and etc. The proposed method let a free point make two dimensional random movement with reflective boundary on feature vector, took the difference and distance of two adjacent points as their movement possibility, and then the equilibrium distribution of Markov chain is taken as saliency values (the situation of combination algorithm is similar). This method can improve the reliability and calculation efficiency of attention fixation determination.Based on the improved methods above, a new spatial attention information extracting algorithm based on contrast sensitivity and Markov chain was proposed. It provided spatial attention foundation for the application of spatial attention algorithm in scene classification and building spatial and object attention cooperative work model. The area under receiver operating characteristic curve and average of algorithm cost time demonstrated that it is an effective spatial attention information extracting algorithm.(5) A scene classification model based on Markov Chain was proposed with the application of spatial attention information extracting algorithm based on contrast sensitivity and Markov chain in scene classification. The model obtained the Gist vector and reduced its dimension with the spatial attention saliency computation method to guide scene classification. Compared with the Gist model proposed by Siagian and etc, our model's contribution is that the model based on the physiological structure of biological visual system, transforms input image into retina image with the retinal transformation method based on contrast sensitivity before feature vector extraction, the equilibrium distribution of Markov chains which are defined on feature vectors are used to extract Gist vector.(6) According to the working principle of biological visual system, a computer model of visual attention based on cooperative work between spatial and object attention was built based on the directional object visual attention algorithm and the spatial attention information extracting algorithm based on contrast sensitivity and Markov chain proposed.The proposed model transforms visual image into retina image first, then carries out spatial attention and object attention according to the task in control center, and records fixation point path to support other cognitive actions finally. Each module and the whole model had been test, and experimental results indicate that the performance of spatial attention and object attention can be improved and it can realize the cooperative work between object and spatial attention with the proposed model.