Research And Comparison Of Pattern Recognition Based On Image Feature

A large number of digital images need to be analyzed and processedwith the increasing power of computing and the fast development ofinternet. It is an important application that images are classified byimage features. Therefore, pattern recognition is one of the most activefields. There are three parts for a pattern recognition system:pre-processing, feature extraction and pattern classification. The papermainly studies problem of image recognition, which can be summarized asfollows:For pre-process, at first, the color images were converted to grayimage. Secondly, we applied median filtering to image de-noising. At last,we applied EM algorithm to image segmentation.For feature extraction, this paper mainly studied the following fourkinds of feature extraction. First, the principles and methods of momentinvariant were discussed, and a new algorithm for moment invariant waspresented. Second, the continuous Fourier descriptor method was appliedto feature extraction for making up defect of discrete Fourier descriptormethod. Third, because the dimensions of images are very high, recognizingthe high dimensional images increases the cost of the system. The paperadopted the principal component analysis (PCA) as feature extractionmethod. It reduced the dimensions of images. Fourth, independentcomponent analysis (ICA) is another useful tool for high dimensional dataanalyzing after PCA. ICA method made up PCA defect about the lack of highorder statistics of image data. The paper adopted FastICA as featureextraction method. It reduces the number of iterations and speed up theconvergence of the system.For classification, the principle of BP neural network and RBF networkwas discussed, and their respective advantages and disadvantages of two types of networks are summarized. Afterwards, a RBF-BP combinationnetwork was applied to classification. The new combination network notonly has better generalization performance of BP neural network, but alsoprovides preferable convergence speed of RBF network.Feature extraction experiment results show that moment invariantsmethod and continuous Fourier descriptor method as feature extractionhave the same recognition rate. However, PCA method and ICA method asfeature extraction have higher recognition rate. Compared with PCA method,ICA method has higher recognition rate than PCA method. Classificationexperiment results show that RBF-BP combination network not only have thehighest recognition rate among three kinds of networks, but also canovercome the shortcomings of classification methods based on single BPand RBF network.