The content-based image retrieval is an important research point in recent years. It can help finding images from the database exactly and quickly. Based on the image texture, this paper presents a new classification algorithm in which the wavelet analysis is performed, after the careful analysis, designs the features reasonably. Then, according to these features, the similarities of the textures are computed and the classification result is obtained. In the experiment, 100 images are searched and 75% of the results are correct. In order to further prove the robustness of the algorithm, 35 kinds of texture images are rotated to get 35 new images, then these 70 images are searched, the correct rate is 64%. The experiments show that the algorithm is effective and most of the textures can be correctly recognized by using it. Compared with other algorithms, it not only has high correct rate, but also is robust with respect to the results of the rotated images. According to the analysis of the distance between different textures, the algorithm could reflect the similarity of similar textures and meanwhile, does not necessarily require that the images are completely the same in size.