Visual Attention And Cognitive-behavioral Model And Its Applications

Vision is a mian manner of human perception.Since visual images have abundant information and high dimensions,attention selection mechanism plays a key role in the initial stage of visual perception.Attention selection helps us to analyze and select interesting areas and proto-objects in vision scenes,which leads to a favorable application foreground.Whereas,existing visual attention models have high computational complixity and need careful parameter selection that make them hard to meet the requriment of real-time projects such as image and video quality assessment.Choice behavior,similar to visual perception,is another basic skill for beings.It is a high level information process of brain.Reinforcement learning algorithm in artificial intelligence is able to simulate the learning process of beings' choice bahavior by encouraging and punitive signals and unceasing trial-and-error interactions with a dynamic environment to achieve an optimized strategy.However,the traditional reinforcement learning method based on look-up table can hardly deal with high dimensions or be applied in the learning of visual information.Considering the above limitations,we do some research in this thesis.Our contributions are shown as follows:1.we propose a simple and quick method to simulate attention selection in frequency domain called as PFT or PQFT.We set the amplitude of each spectrum to the same constant and keep the phase spectrum,then transfer to spatial domain.After filtering,the saliency map can be obtained.This algorithm increases the computing speed substantially,and can extract more accurate saliency features of a given image comparedwith existing visual attention methods.2.we propose saliency-based image and video quality assessment methods.We extracte the saliency maps from reference images or video frames by our PQFT method.Then we treat the saliency map as weights to adjust the original image quality assessment criteria.In video, both inter-frame and intra-frame weighted quality assessment criteria is proposed according to saliency map.This saliency-based strategy improves the performance of quality assessment significantly and is more similer to subjective assessment.3.We propose a neural network model to simulate the choice behavior of insects facing visual cues based on biological facts of Guo's group.The proposed model introduces value system and a reinforcement learning algorithm based on the value system.It can learn input visual images under the reward or punishment and establish a value system,which simulates choice behavior based on dopamine and mushroom body circuit in the brain of insects.The proposed model can simulate the development of value system and choice behavior of drosophila.The simulating curves accord with biological experiments.This model provides a basis for the further application on robot auto-control system facing visual cues.4.We combine visual attention and reinforcement learning algorithms and apply them on a 3d simulation platform to realize the vision-based autonomous driving of intelligent vehicle in unknown environment by C++ launguge.