Image Classification Based On The Low-level Feature And SVM

In recent years, with the development of computer technology and increasement ofimage database, how to extract visual information from image is more and moregetting the attention of people. Classification and retrieval for image database havebecome an important technology to get image information. And image classificationcan reduce the scope of image database, and made the retrieval effect more obvious,so the image classification has a very importantly practical value. Image classificationresearch focuses on two issues: the feature extraction of image and learning machine.The texture feature and color feature are the important low-level feature for images.Gabor wavelets transform is developed on the basis of Fourier analysis, which is agood signal and image processing method. Support vector machine (SVM) is a kindof learning Machine based on statistical theory. It can change the original nonlinearinseparable problem into a linear separable feature space for classification study. Thework of this paper is mainly concentrated on two the last aspects:In first place, a image classification method based on color Gabor texture featuresand SVM is proposed, which avoids the loss of color texture in feature extraction part.Gabor wavelet has good performance for gray image, but for natural color image it’smissing part of the color information. So we extract the texture feature in the image ofRGB three-channel and form the pseudo color image color texture feature used forimage classification. This way can keep part of the color information of color image.We use the Corel1000as the image database, to verify the proposed method in thispaper has good classification effect on natural images.In the second place, we presented a kind of intelligent image classification based onfeature fusion and the SVM algorithm. For a natural image, the single feature alwaysdoesn’t represent all of the image information, but the multiple features can betterrepresent image information. In this paper we use the low-level features of eachindividual to train a SVM classifier, and then we use particle swarm optimization (PSO) to determine the weight of each classifier and then used for classifying colorimages. Experiments based on natural image Corel1000prove the proposedalgorithm for natural image has good classification effect.