Research Of Motion Detection And Integration Based On Biological Vision Mechanisms

Detecting and processing motion information in image ha ve been widely applied inengineering and military. With the development of IT technology, many effectivealgorithms for processing digital image have been developed. However, compared with theoutstanding performance of biological visual system in processing complex images, thereare still many inadequacies. By now, one of the most exciting and open research frontiersin neuroscience is that of seeking to understand the functional roles of the layers ofcerebral cortex. So creating a bionic brain based on mechanisms of biological visionsystem is a very promising field.The main object of this study is the mechanisms of motion pathway of the primatevisual system, involving the retina, LGN (lateral geniculate nucleus), cortical visual areas ofV1, MT and MST. By tracking and in-depth analysis findings of modern biology inprimate visual system, we discussed mechanisms of detecting motion in retina, pathwaysof motion information, properties of neurons which are involved in motion processing andtheir topological relations, mechanisms of integrating motion information in cortexes,functions of feedback projection and functions of lateral horizontal connections in cortexes.This paper develops a neural network model for detecting local motion and integratinglocal motion from large scale to get the global motion, and emulates the model with C++Builder. Motion pathway of primate visual system was systematic analyzed withmathematical tools of Topology and Category, and the related mathematical models wasdeveloped.From a theoretical point of view, local motion detection mechanisms have to satisfy atleast three requirements in order to signal motion in a directionally selective way. That is, amovement detector has to be asymmetrical and needs at least two inputs from differentlocations, which interact in a non-linear way. This paper treats response of bipolar cellsfrom two different locations as input to visual system, utilizes starburst amacrine cell which have lateral inhibitive connections with nearby starburst amacrine cells to facilitatedirection-selective ganglion cell asymmetrically and to inhibit local direction-selectiveganglion cell. We systematically analyzed the response of this model in many kinds ofconditions.By analyzing findings of physiology and anatomy in primate visual system, we listedthe organization of cortexes and connections between them systematically. We found thatvisual information is not always feed-forward through dorsal pathway from low levelcortex to high level cortex when studying information transmitting in motion pathway. Forexample, compared with10％retinal afferents,30％of the synaptic input to LGN relaycells is from feedback connection with cells in layer6of V1which regulate response ofLGN relay cells. There are also massive of horizontal connections between differentpathways, such as connections between V4and MT. Visual system utilizes globalinformation retroinjected from high level cortexes to guide further processing in low levelcortexes and carry out information interchange between different pathways with thisfeedback and horizontal connections.By restrictions of the scale of receptive field, low level neurons, such as that in retina,LGN and V1are confronted with aperture problem which means that motion of afeatureless line seen behind a circular aperture is perceptually ambiguous, for any realdirection of motion, the perceived direction is perpendicular to the orientation of the line,called the normal component of motion. Thus, the visual system needs to use the relativelyfew unambiguous motion signals arising from image features to constrain the morenumerous ambiguous signals from contour interiors. Global motion is succeeded perceivedby combination of mechanisms of lateral inhibition, coupling and feedback. And problemof feature detecting and matching is also avoided by this model.At last, the validity of the model is verified by simulating with programme built withC++Builder. Effect of MT feedback connection on V1is also simulated, which proved thatregulation of feedback connection can notability inhibit ambiguous motion information.