Research On Natural Image Processing Based On Theory Of Sparse Coding

Biological experiments show that response properties of visual cortex have been considered to be sparse neuro-representation. Sparse coding has its origin in the study of visual neural network; it is a neural network method for finding a representation of multidimensional data in which each of the components of the representation is only rarely significantly active. Sparse code theory establishes a scientific quantitative link between the information processing mechanisms of visual neurons and the statistics of input visual stimuli, and provides an efficient tool to understand the neural information processing mechanisms. It has been applied in blind source separation, speech signal separation, image feature extraction; natural image denoising and pattern recognition. It has achieved many fruits and has important practical value.The sparse coding of natural image is an artificial neural network method, which can model the receptive fields of simple cells in the mammalian primary visual cortex in brain. The encoding realization for this method only depends on the statistical properties of natural perceptive information, regardless of the inherent properties of input. It is a self-adaptive signal statistical method.This thesis presents the theories and algorithms of Sparse Coding (SC) and explores its applications to the natural image feature extraction, image denoising. The main work of the author focuses on the following aspects.We give an overview of the status of research on SC. Introduce some SC-related knowledge: statistics and information theory. Introduce the statistical properties of natural scenes and its linear sparse coding.We study the typical algorithms of SC that have more influence at present and write program to implement these algorithms. In the experiments we combine standard sparse coding algorithm with Gabor filter to find an efficient coding of natural image. In addition we use sparse code shrinkage on coefficients to reduce the Gaussian noise. All experimental results show that sparse coding method that makes full use of natural image data in the high-order statistics gets an excellent analysis results.We also compare SC method with some traditional methods: PCA and Winner filter algorithm. The comparisons show SC method has better performance in image feature extraction and denoising.Research shows that natural image processing using SC method, which compares to traditional image processing methods based on theories of digital information processing and probability statistics, has unique advantages. It provides us a novel approach based on information processing for image processing.