Research And Application In Conceptual Mapping Method And Learning Algorithm Of CMAC Neural Network

As the popularity of on-line categorizing shopping at Internet and color printing, color images as the information carriers have been paid more and more attention. When a color image goes through several operation platforms or image devices, it always losses some colors. It's core contents of color matching how to get the image with same color information when it turns out with color printer or display. And it's also a key technical problem in the field of color information process.CLUT (Color Look-Up Table) defined in ICC Profile is the international standard solution of the problem of color matching. But CLUT only contains part of color transform information due to limited memory, so interpolation algorithm is needed to calculate the data that weren't contained in the Table. But the color matching refers to high-dimensional non-linear transform, so neural networks that have the advantages of non-linear functional approximation and fast learning speed and parallel process is suitable for the interpolation algorithm. This paper applies CMAC (Cerebellar Model Articulation Controller) neural network as interpolation method for Multi-Dimension CLUT and has gotten good results.CMAC neural network adopts lookup table (LUT) technique and has the advantages of fast learning, Local generalization and easily hardware implemented. It has been applied in the fields of robot control, pattern recognition, signal processing and color matching which need large computation and fast learning speed.However, there are some short-comings of CMAC neural network, such as low learning accuracy, noise caused by address conflict and no global generalization ability which limit the development and application of CMAC in some field. It's very important and necessary to do some research on how to improve the algorithm of CMAC neural network.The main research contents of this paper include conceptual mapping methods and improved learning algorithm. A new conceptual mapping method based on main (anti) sub diagonal has proposed in the paper. It can get good learning accuracy and generalization ability even in limited memory. Meanwhile, B-Spline CMAC and Fuzzy CMAC neural network have been improved combined with new conceptual mapping method. Simulation results show that the integrated performance of both kinds of CMAC structure has improved. On the other side, learning algorithm based on Credit Assignment has taken the place of LMS (Least Mean Square) as the CMAC learning algorithm.