Research On Image Representation Model Based On Non - Classical Feeling Field And Its Application

Image is an important source of obtain information from outside world, and is also a key medium of information communication. How to realize an effective and practical method of image representation, it is play an important role in understanding image processing deeply and provide in-depth support to the further image processing. The efficient image representation model has some important function such as how to get the image coding and computer vision processing mechanism more efficient further. Currently, the prevalent image processing method is measured in pixels. These theoretical approaches from engineering view are faced with great amounts of data, low processing efficiency, high hardware requirement and cognitive impairment in complicated scene. By contrast, human vision system can be easily recognizing object with all kinds of features and scales. As a result, with the aid of neural mechanism and cognitive psychology theory, it is entirely possible that find a new approach to solve the bottleneck problem that restricts the image cognitive in computer vision at present. The research has its foundations in Major Project of Chinese National Programs for Fundamental Research and Development (973Program2010CB327900) and perform research work as follows:First, a hierarchical neural computation model with feedback using non-classical receptive field (nCRF) of retinal ganglion cell (GC) is established. We utilize the dynamic adjustment characteristics of RF according to image stimulus, set up a new image representation unit that takes on image texture analysis and extraction. The computational model have top-down control function, by reference to attribute of image stimulus to realize neurons’ function of integration and active selection.Second, the GC-array, it is result of RF integration according to the color or texture change of image, can represent the image efficient. A large size RF can represent the certain closest color region of image. On the contrary, a drastic color change region can be represented by a small size RF. To test the representation fidelity using GC-array, some experiments demonstrate the relationship between image statistical characteristics and RF quantity. The final result shows that the GC-array can reflect the image properties faithfully.Finally, the aim of image representation is to improve the further high level of image processing efficiency. Some experiments such as feature matching, image segmentation demonstrated that the new image representation (GC-array) enhance the performance of the algorithm greatly. We design a contour detection model using nCRF mechanism. The idea is to create a new algorithm that integrate inhibition and disinhibtion characteristics of nCRF together, utilizing the multi-scale information of images, and do a reasonably good job at "weak contour" or "much noise" in contour detection. A quantitative analysis method has proved that nCRF enhance the performance of contour detection algorithm.