Aiming at the problem of the recognition accuracy of deep belief network is not too high, the concept of difference measure based on fuzzy membership function is introduced, a new image classification method, or D-DBN for short, of deep belief network based on the sparse difference is proposed in this paper, and its application on insulator identification is put forward. Because the difference theory has the advantage of widening the low gray areas, and reducing the high gray areas, it more consistent with characteristics of human vision. At first, the images were changed from gray feature matrices to the difference feature matrices, then the difference matrices were made mean, normalization and sparse . Secondly, the difference features were trained by DBN to learn more intrinsic characteristics of data, so as to achieve the aim of improving recognition performance. The recognition results on MNIST and SVHN database with different sample sizes and different network structures demonstrate that, comparing with the traditional DBN method and other improved methods, the proposed method achieves better recognition performance. At last, a method of D-DBN was applied in the fault identification of insulators.