Research On Image Recognition Based On Moment Feature

During the recent 20 years, Image recognition has developed to be a technique of Pattern Recognition, which achieves human ability of visual cognition and comprehension by modern information processing and computer technology. Image recognition principally completes images classification according to feature extraction, diffusely applied to kinds of fields including character recognition, fingerprint identification, remote sensing, medical diagnosis, industrial products test, satellite photos and aerial photographs interpretation and so on.As extracting image features, we make use of invariable moments consisted of Hu , Zernike and wavelet moments to achieve target images'features ,which possess invariability of translation, rotation and proportion.Hu and Zernike moments are calculated in whole image-space and get the global characters ,as a result they aren't classified easily. Especially under the environment of limited amount of samples, it's more important to select the best and most expressive feature. Wavelet moment based on wavelet transform is able to acquire images'global and partial feature at one time and more propitious to recognize images with similar shape and noise.In respect of pattern classification, BP neural net obtains the classifying result just passable because of its defects that blindly designing the framework, steepest descent in slowly converging , easily immerging in partial minimum frequently and discriminating astatically. To deal with the flaws of BP net, we improve the network topology for wavelet neural net however which still with such defects as complex initialization of parameter and huge amount of calculation in the case of more dimensions.Support Vector Machine(SVM)is an burgeoning discipline based on statistical Learning Theory (STL) of the VC dimension (Vapnik Chervonenkis Dimension) theory and Structural Risk Minimization(SRM) principle , which possesses better analytical and problem solving skills to solve the local minimum point, high dimensional, nonlinear and other practical problems.Method of image recognition based on statistics can achieve fine performance only if it's provided with large numbers of samples. In practice however, sometimes it's impossible to obtain so many samples,which may results in the poor recognition-performance because lacking of information. Consequently in the paper an arithmetic that combines wavelet moment with Support Vector Machine is established.To verify the effectiveness of the algorithm in the paper, we operate the computer simulation experiment on the platform of Windows Vista operating system and MATLAB 7.8.0 , and take five kinds of tank ,three for each kind, a total of only fifteen standard binary tank images as small samples for training.In the case of noise and noise-free respectively, there are 300 tank images which we extract their features and classify them. Experimental results demonstrate that the arithmetic which combines wavelet moments and SVM is superior to others on recognition efficiency in the case of small sample.