Dynamic Image Analysis Research Based On Biological Vision Principle

Dynamic image analysis is a highly active field. The traditional computer vision theory encounters lots of difficulties in dealing with motion issues. This paper made a detailed research on biological vision principles and proposed a neural model of motion perception for natural image sequences based on the recent findings. Also, the simulations and analysis were concerned.The visual cortex related to motion and direction perception as we know so far involves area V1, V2, V3, MT and MST, where the path of V1-V2 is static image processing system and the path of MT-MST is motion processing system. The Formotion BCS model proposed by Grossberg is a successful motion perception model based on these findings. However, it emphasizes on the analysis of visual psychological phenomena and cannot be used for the motion perception in real scenes.The article studied the basic features of motion perception and the special requirements for natural dynamic image processing and synthesized several neural models to put forward a new parallel processing mechanism of dynamic image processing in real scenes. Firstly, the SBCS model was improved to make pretreatments to various image conditions like luminance equilibrium and to compute the static image boundaries; secondly, a figure-ground separation mechanism was designed to detect moving targets. It could not only predigest the processing conditions but also make good solutions to the occluding problem; finally, the key model was proposed for motion direction perception. The fist stage of this part was transient cells level which was used to detect the directional changes. But the detected moving directions were spurious because of the aperture problem. Then signals went through the short-range filter and competition levels to extract feature tracking signals. At the stage of MT-MST feedback circuit, the feature propagating mechanism based on long-range directional grouping and attention priming realized the motion capture and obtained the coherent direction perception of the whole moving targets. Additionally, the feedback between static system and motion system further strengthened the amplitude of feature signals.Various image sequences were chosen to be simulated by the model. The results indicated that model was exactly computed and was available to kinds of conditions, which also showed the model accorded with the motion processing mechanism of the biological visual system.In the end, the paper summarized the research work and prospected the future studies.