Based On Visual Characteristics And Artificial Neural Network Image Compression

Image compression is mainly dedicated to eliminate the information redundancy such as coding redundancy, interpixel redundancy, phychovisual redundancy. In the past few decades, with the fast development in image compression research, a lot of efficient algorithms have been proposed and a few international standards have been made, such as JPEG. JPEG2000. MPEG,H.26X and so on. Among the three kinds of information redundancy, algorithms about eliminating coding redundancy, interpixel redundancy in image compression have come to maturity, so limitation has been reached in the traditional image compression methods based on Shannon information theory. In order to go a further step in image compression, two kinds of research can be made:one is to using the characteristics of human visual system as eyes are the terminal "consumers" to the images and researches based human visual characteristics has more and more become the "hotspot" in image compression field. The other is to exploit new compression tools more intelligent algorithms. Artificial neural network possesses large potential to image compression because of its excellent performance for information processing and researches in this field are in the primary stage in the current period.We have do some research at the two direction in the thesis and our contributions are shown as the following:1. The visual efficiency of image compression depends directly on the amount of visually significant information retained. Psychophysical studies show that the human visual system is sensitive differently to edge, smooth and texture regions, and the wavelet coefficients within different subbands. Based on these characteristics of human visual system, this paper proposes a new technique that implements wavelet-based image compression at low bit rates. By using diverse quantization step-sizes in different regions and diverse weights to subbands, wavelet coefficients are reordered and coded according to their perceptual importance. The experimental results show that the proposed scheme performs better in both VIF criteria and subjective quality of the reconstructed images than conventional SPIHT algorithm, especially at low bitrates.2. We proposed a new image compression algorithm based on artificial neural network in the wavelet domain and Layers of the neural network are designed as referring to the visual sensitivity to the subbands. The feature, compression using the image data directly in the traditional algorithms based on BP neural network are remained in this algorithm.3. We proposed another neural network-based image compression algorithm that images can seen as the function of axis and we can use the BP network to train to "mimic" the image and transmission the weights which is far less than the image data to achieve image compression. In the neural network, the axis of the data as the inputs and the data as the "teacher". Images can be decoded using the weights gained by training. Comparing to other neural network-based methods which directly use the image data to do compression, the algorithm proposed have improved capability of data compression effectively...