Abstract: As wavelet cannot well express the edge curve characteristics of coal-rock images,leading to a low recogni- tion accuracy,a method based on curvelet transform was proposed to have a sparse representation of coal-rock image edges. The method used the curvelet transform to decompose images into curvelet coefficients in different scales. The low-pass coarse coefficients were preserved and then Gaussian random matrices were used to measure the high-pass co- efficients in order to realize a dimensionality reduction based on the compressed sensing theory. The feature vectors for coal-rock images were created by concatenating the low-pass coarse coefficients and high-pass coefficients after dimen- sionality reduction. Finally,classification and identification were carried by support vector machine. Experimental re- sults showed that the features extracted by curvelet decomposition could effectively express the curve features of coal- rock image edges. The classification accuracy of the proposed method reached 93. 75% . And it improved the classifica- tion accuracy of 4. 37% than Haar wavelet method. The feature vectors extracted by the proposed dimensionality re- duction method were more advantageous to the classification and recognition of coal and rock images than the linear di- mensionality reduction methods.