Image Segmentation Based On Genetic Algorithm And Neural Network

Image segmentation is a process of separating an image into several un-lappedregions, in each which all pixels have similarity in intensity, color or texture, etc. Imagesegmentation, as the basis of image analysis and understanding, is one of the mostdifficult problems in computer vision.BP neural network has the capacity of parallel computing, distributed saving,self-studying, fault-to-learnt and nonlinear function approximating. So it's widely usedin the classification area. Recently, there are vast literatures about image segmentationbased on BP neural networks in the image processing area. Babaguchi N use BP neuralnetwork to segment image. It regards gray feature as input, and the thresholding ofimage segmentation as the output. But BP algorithm has some unavoidable defects suchas slow speed in training (especially for large training samples), liable to get into localminimum, and so on. Then some people present some methods to improve BP algorithm,such as improve the searching pace, and therefore improve the training speed, but thisapproach also can hardly avoid the BP algorithm getting into local minimum.Considering of the ability of Genetic Algorithm (GA)'s global searching, someresearchers use GA to optimize BP neural network to avoid it getting into localminimum and obtain effective result.After analysis of the feasibility of algorithm combined BP with genetic algorithm, wepresent a approach to combine modified GA algorithm with BP neural network (shortfor MGA-BP), give the details of this approach and a concrete implementation.According to the actual situation, we adopt real code, intermediate crossover andGaussian mutation to optimize BP neural network in this implementation. Throughemulation, we have compared the training results between modified GA-BP (MGA-BP)algorithm and separate BP algorithm, the results show that MGA-BP algorithm cansearch for the best result rapidly, and also avoid BP neural network getting into localminimum. At last, we use this algorithm to segment the medical images with the bestthreshold found by the trained MGA-BP neural network. The result of emulation showsthat the MGA-BP achieves satisfactory effect on these medical image segmentation, andtherefore has a good competitive power on the medical image segmentation.