Image Recognition Method Based On Artificial Neural Network Research

The image recognition is developing in the last 20 years.The image recognition is become the main contest of the classification and description in some objects or process(regarding as the Image).This dissertation emphasizes on the research of air target recognition, which can be divided into two parts including preprocessing and post classification of the images. This paper is structured as follows.At first, the source images are being preprocessed. Robert, Prewitt and Gauss-Laplacian operators are used to detect edges. And the images are iteration thresholding segmented and binaried. Then, Hu moment feature is extracted, normalized and modified. At last, moment feature, which is the input of artificial neural network, neural network ensemble and Support Vector Machines (SVM) respectively, is used to recognize the category of the aircraft.In this dissertation, the structure of BP network and classical back-propagation learning algorithm are systematically analyzed. An improved BP learning principle, based on the standard BP algorithm, is presented, which stressed on the adjustment of learning rate and vector factor. The parameters of learning rate and vector factor are adjusted to improve learning speed in dynamic and self-adjusted way.In this dissertation, an improved Genetic-BP algorithm combining of GA and BP algorithm is proposed, which describes the global optimization of multilayer feed-forward neural network as the heuristic genetic search problem. Using GA twice, the new algorithm can correspond the dynamic adjustment of search space parameter with the dynamic training character of neural network, which not only overcomes the optimization of blindness, but also avoids the occurrence of local convergence.Neural network ensemble including Bagging and Boosting is applied to classification and recognition of the images. Experimental results show neural network ensemble has better generalization than individual neural network.With the feature of small sample of the used data, SVM is used for target image recognition and classification. Three methods of multi-class and three kernel functions SVM are tested. And better experimental results are achieved than that of neural network and neural network ensemble.