A quality grading method based on sparse representation was proposed to identify the varieties of tobacco quality. The images of 17 different qualities of tobacco were taken as objects. Ten images of each variety were selected randomly as training samples. The colors, morphological and textural characters of these images were extracted for making up the dictionary of sparse representation. The projection of the test image on the dictionary was calculated. The minimum projection error was regarded as the certain kind of tobacco. The result of the proposed method was compared with basic pursuit algorithm, neural network, SVM and fuzzy processing. The identification accuracy of training samples was 100% and the overall one was 95.7%.