Study On The Feature Modulation And Selection Strategies For Modeling Visual Attention Mechanisms

Human vision is the most important means of obtaining information from the outside world, and the research of visual information is the important issue of scientific research. From the invention of computer, people are dreaming that one day computer could understand and change the world in a way human beings do. But, nowadays the ability of computer vision is far below the ability of human vision. For this reason, scientists have proposed many methods which are consistent with human visual processing characteristics in order to improve the capacity of computer vision. The computational model of visual attention mechanism (VAM) is an image processing method rising in recent years.The main work and innovation of this thesis are summarized as follows:Former studies have only focused on the bottom-up image information, however, the integration of bottom-up (data-driven) image cues and perceptual properties of the human visual system is very important for the selection of salient locations. In this thesis, a new data-driven VAM computation model matching the visual perception is put forward, in which the brightness-adaptation characteristics (also called the luminance threshold effect) of human vision is applied. In more detail, the primary-visual features (e.g., the intensity and orientation) are adjusted using the sensitive value of brightness in the local scene of the feature map, and then a saliency map is combined based on the perception. Experiments show that the prediction results using the proposed model are the same as those of human beings.One of the great challenges of modeling task-driven VAM is to find out the optimal top-down effects on feature maps so as to maximize the detection speed. The ratio of the mean salience-value between targets and disruptors had been used as the weights of the feature maps during the searching period. This algorithm only works when the salience value is uniformly distributed in the feature map, which is not common in natural scenes. A novel optimal feature-weight regulation strategy based on the stimulation intensity ratio between targets and disruptors is proposed in this thesis, which can maximize the relative salience of the goal. The stimulation intensity is determined by two factors, i.e., the mean activity coefficient and the cumulative summation of salience. A pruning method (i.e., a feature-map subclass having a weight greater than the given threshold is extracted for the final feature-map combination) is proposed to reduce the computational cost during the search process. Experiments with artificial and natural scenes show that the speed and accuracy of searching objects can be improved using the proposed optimal feature-weight regulation strategy and the pruning method.