The Study Of Texture Feature Extraction Of Gastric Epithelial Cell Microscopic Image Based On Sparse Coding

When we analyze and identify microscopic images of gastric epithelial cells using computer, one of the key technical aspects is the feature extraction. What kinds of features should be extracted make sense in the recognition algorithm design and the practical recognition results. As natural images, microscopic images of gastric epithelial cells have the following characters: (1) the types, structures, shape, sparse level, and arrangement of the cell will be different and complex for the irregular shape of tissues, organs and the difference of the different cell classes; (2) its high-level statistical properties characterizes as a non-Gaussian distribution; (3) it contain lots of redundant information. Because of the limitations of linear methods, when we use the general linear methods to extract features, the extracted features aren't effective. Therefore a new nonlinear feature extraction method is presented in this paper. This method which can use of the image of their own characteristics and automatically extract the invariant features to reduce the complexity of the image is based on the principle of sparse coding of visual systems, can improve the classification accuracy.We have been able to get the texture features to divide the images to be recognition into three types: normal, hyperplasia, cancer and achieved good recognition results by using the new method to recognize the microscopic images of gastric epithelial cells. It demonstrates that this new method is effective.In this paper, we have done the following tasks:Firstly, a visual sensor network which contains a sensor layer, a bipolar cell layer and a ganglion cell layer was build by imitating the human visual system. The network can get the contrast enhancement and edge highlight information from the gray level information of microscopic images of gastric epithelial cells, and further encode this information into groups of ganglion cells excited information. Through such conversion by visual sensor network, the original information has become nonlinear compression information, which is reducing complexity and eliminating certain redundant, and providing favorable conditions for feature extraction and further processing.Secondly, a visual cortex neural network with mini-columns as its functional unit was constructed. The network automatically extracts two types of information: (1) the information of the changes in size and trends of the input information in a period of time. The visual cortex network can extract the information on changes in a certain period of time; (2) the information of the probability of an input or output model repeated. Both the information stored in specific neurons of micro-columns after extracted. The sparsest respond of these specific neurons to the input image was optimized by clone selection algorithm. Meanwhile, this algorithm also makes the response of the specific neurons to input adaptive and robust. The micro-columns which contain these specific neurons are further encoded image features by the response of dynamic population. These image feature coding has good sparse and nonlinear characteristics, can be effectively expressed natural images and its properties.Finally, the complete texture primitives were extracted on the basis of the dynamic response coding of micro-column groups in the above. These texture primitives used for texture representation and combinations.