Image compression using artificial neural network
This thesis consists of two parts. The first part of the research focuses on theoretical issues and algorithm enhancement. First, the linear and nonlinear neural networks were analyzed. Theoretical analysis and a description of the advantages and disadvantages of linear and nonlinear neural networks were provided. Secondly, an analysis was provided on how to choose the input vector size. The result obtained using a neural network with a large input vector size is clearly better than that obtained using a neural network with a regular (8x8) input vector size and the compression ratio is larger as well. Thirdly, the theory of image compression using preprocessing classification was presented. With this theory, the input vectors are preprocessed and then they are classified into different groups with the variance criteria. A distinct neural network is used for each group. Using this method, a higher compression ratio and better restored image can be obtained. The cost is more complex computation and consequently more time for the preprocessing classification procedure. In the second part of this thesis, the hardware characteristic of Intel 80170NX neural network chip is analyzed, and image compression theory is implemented in a 80170NX chip. I first describe how the neural network idea, such as parallel distributed processing, adaptive weight modification, and synapse multiplication, are implemented in hardware. Then the characteristics of the 80170NX chip are analyzed and the experimentally obtained curve of the sigmoid function of the 80170NX chip is presented and its affect on the 80170NX neural network system is analyzed. Finally, the results obtained using the neural network embedded in the 80170NX chip are presented. Although the result is not as sharp as that obtained from simulation, it still looks good. The processing time using 80170NX is very fast. And this shows the potential of commercial usage of image compression using neural network.