In order to increase the accuracy classified by image texture to remove plenty of useless information in wood hyper-spectral images, the texture classification algorithms based on the multifractal theory were used in this paper. Firstly, the most representative ten bands were screened out using different Feature selection algorithms, and then the multi-fractal curves of the selected bands were obtained according to different function density images. The multi-fractal curves were averaged, which could represent the texture feature of certain sample. Finally, the curves were classified by the Support Vector Machine (SVM) and BP neural network classifier. The result shows that the bands screened by K-L divergence are superior to that screened by adaptive band selection (ABS), the image’s texture features extracted by multi-fractal algorithm are better than that extracted by the gray level co-occurrence matrix(GLCM), and the classification accuracy and classification speed of SVM have an advantage over that used BP neural network. It can be concluded that the integrating K-L divergence, multi-fractal with SVM algorithm can dramatically increase the recognition classification, which can reach 97.91% in our work.