Feature Extraction And Auto-recognition Research On BP Networks Of Cell Image

Nowdays, digital image processing is widely used in biomedicine engineering field. The aoto-recognition of digital image becomes the research focus consequently. This paper extracts 27-dimension features on the base of preprocessing of digital image. These features include moment feature and texture feature, and this paper proposes a fast and precise algorithm of moment feature.On the base of compare of the traditional statistics pattern recognition and ANN, this paper constructs feedforward network with back propagation to auto-recognize and train the body-fluid cells. By emulation with Matlab, this paper verifies three different methods of parameter optimize:GDM, RPROP, Levenberg-Marquardt method '.Different network architecture, error performance function and trainning time are selected with each method to find the optimal combination architecture.This paper establishes a general recognition module, and this module has great adaptability, and it is intense integration. This module is concluded in a class of vc++, the input dimension , the numbers of hidden layer neuron and output layer neuron are controlled by parameter. Users can explore the network only by adding corresponding files to the project and give the parameters.This paper sets worker thread and user interface thread to control trainning process at any moment. In the course of trainning , one can intervene with the trainning process , and can stop the process and change the parameter to achieve the optimum results. The trained network would be saved specified file after trainning. When a pattern is to be recognized, the saved file will be transferred. If a pattern is unable to be recognized, it can be added to the samplecollection as a sample, and the sample collection is expanded consequently.Anything involved in this paper is finished on the platfonn of VC++, the research result is verified by the recognition of real cell images.