Adaptive self-excitation groups in visual curve integration

To enhance visual curves by neural models, many authors have used a self-excitation (S-E) neuron group inside which neurons are excitedly interacted. Typically, an S-E group is immovably arranged as a subnetwork in which S-E actions are done through the direct excitatory connections between the neurons. These models are often only fit for simple and fixed input patterns and have no ability to dynamically self-organize S-E groups on the basis of external inputs. This article presents an adaptive S-E model that can dynamically self-organize various S-E groups according to actual inputs. The significant merit of our model is that the organization of each S-E group is temporally separated from the others. As a result, a local neural circuit can be shared by multiple related S-E groups. This model consists of three parts, heuristic curve-searching neural structure, time filter, and accumulation representation. The curve search is realized by random walks of neural impulses. An S-E group is temporarily constructed via the instruction of a continuous search trajectory. Different S-E groups exist temporally separately. S-E action is dependent on synchronous impulses that are implemented by time filter. The repetitive searches as a statistical method are to accumulate curve stimuli and cut down the effects of noise. Finally, many experimental results show that our dynamic S-E model can work well in noisy images.