Twenty-six color, shape and texture features were extracted from seven kinds of cucumber disease leaf. It was found that the sparsity coefficients for different features had similar structures when they were sparse represented by the same training set. By introducing the joint sparse model to construct the cost equation, thus the regularity was summarized in mathematics. The joint sparse coefficients were solved by using the accelerated proximal gradient method. Finally, disease recognition was realized by means of reconstruction error. Experiments demonstrated that the correct recognition rate of this algorithm reaches 90.67%, which is 5.7% higher than that of the sparse representation classification algorithm, and the computational consumption time is 7.5 s, shortening 4.3 s than that of the sparse representation classification algorithm.