Image Segmentation And Shift Of Visual Attended Regions Based On Neurodynmaics

The paper analyzes single neural oscillators, proposes and implements biological neural networks(BNN) based on Wilson Cowan oscillator or relaxation oscillator. The composed BNN is used for image segmentation and shift of visual attended regions. The paper introduces the concept of lateral potential, which contributes to the shift of visual attended regions based on leader-follower mechanism. The research fruits are meaningful to discover how human nervous system react facing light stimulation, simulate how human brain process visual information.The paper includes three parts:(1)Analyzing Wilson Cowan neural oscillator and relaxation oscillator from aspect of differential equations. By analyzing single neural oscillator, we study the condition of oscillation and lay solid foundation for composing biological neural networks.(2)Constructing biological neural networks, realize simple image segmentation with computer simulation. The paper composes biological networks with Wilson Cowan oscillator or relaxation oscillator by diffusive connection. The nature of biological neural networks is differential equations. The experiments show that Oscillators limited in the same object starting from random position will achieve synchronization after a while, while oscillators of different objects desynchronize no matter by frequency, phase or amplitude.(3)Implementing shift of visual attended regions after image segmentation following the rule of winner-take-all and inhibition of return. Besides image segmentation based on biological neural networks, the paper adopts the lateral potential and mechanism of leader-follower. After image segmentation by neural dynamics or leader-follower mechanism, saliency associated to all regions are computed with center-surround mechanism. To achieve the shift of visual attended region, the paper sorted saliency values. The method simulates the shift of visual attended regions.The proposed methods are all implemented with computer. Experiments show that these methods are meaningful and useful for modeling biological visual system.