To resolve the problems of high computational complexity and slow training in Convolutional Deep Belief Net, Convolutional Deep Boltzmann Machine (CDBM) is proposed to extract image features. To improve the Convelution Restricted Boltzmann Machine(CRBM), a new training objective function to maximize the probability of intermediate image area is proposed, along with introducing the cross-entropy penalty factor and dropout training. After that, CDBM is designed based on modified CRBM. The mean-pool mechanism is presented to lessen computational complexity and improve the robustness of features for image scaling. The relationship between layers is simplified to extract high-level abstract features. The MNIST handwritten digits database is used to test this new model and the results prove that features extracted by CDBM are more accurate than CDBN. The classification accuracy rate increase at least 0.5%, and training time decrease more than 50%.