A study of the synchronization and desynchronization dynamics of the neural oscillatory network and the applications in scene analysis

Abstract

Theoretical studies of brain functions have asserted that the brain uses the temporal correlation of firing activities of neurons to represent objects in a scene. These studies have been directly supported by many experiments on the visual cortex and other brain regions. In these experiments, the brain exhibited synchronization between the firing of neurons when perceiving a scene. This points to a mechanism of neural oscillation as a representational framework in neural networks. The mechanism has been implemented using a class of neural networks called neural oscillator networks, which it is hoped will provide a physical foundation for the study of human perception as well as an effective computational model for real time scene segmentation. This thesis proposes a neural oscillator network for scene segmentation called a solely excitatory oscillator network (SEON). SEON segments scenes in parallel by synchronizing and desynchronizing oscillators in the networks. The synchronization and desynchronization are very fast. The synchronization speed is also theoretically independent of the network size. This allows SEON to quickly and effectively segment large scenes. A further advantage of SEON is that it is able to reliably segment cluttered scenes into objects. These objects are encoded by the temporal correlation of oscillators. This encoding scheme directly implements the neural oscillation mechanism in human perception. SEON is particularly well suited to very-large-scale-integrated (VLSI) technology, and thus provides an elegant representation for real-time image processing. The performance of the proposed model is evaluated by comparing it with other contemporary methods. Experimental results show that the proposed model can segment scenes faster and more accurately than other methods. The proposed model provides a comprehensive framework for studying human perception and selective attention. It also provides an effective method that is general enough to segment cluttered scenes (both grayscale and color) in real time. Experiments show that the proposed model can be applied to segment cluttered scenes reliably in real time.