A novel image texture exaction algorithm using Independent Component Analysis (ICA) filters for image classification is proposed. Firstly, image patches are selected randomly from images in training set as observation signal. A group of filters (ICA filters) is extracted from the sample texture images using ICA method. And then, ICA filters are evaluated and selected according to the response of the input sample images for the purpose of reducing feature dimension. Finally, response results of test images are gotten by using the filters to the images. Global and local features are extracted from the histogram of the maximum response results. Experiments are completed on the images with 108 texture classes from the Brodatz album. Experimental results show that the proposed texture feature has better classification rate than that of MPEG-7 texture descriptors.