This paper aims to overcome two major defects with the traditional rock image classification framework based on convolutional neural network (CNN), namely, slow training and poor classification accuracy. First, the causes of the two defects were analyzed in details. Through the analysis, the slow training is attributable to the information redundancy in the original image, and the classification error to the lack of differentiation of rock features extracted from the spatial domain. Therefore, the original image was divided into multiple blocks of equal size, and the discrete cosine transform (DCT) was introduced to process each block. After the transform, ten or fifteen frequency coefficients in the upper left corner of the 2D frequency coefficient matrix were retained, and added to the traditional CNN framework for image classification. Experimental results show that the proposed DCT-CNN framework outperformed the CNN framework in training time and classification accuracy.