Community structure is one of the most important features of complex network. Community detection is of great significance in exploring the network structure. Classical clustering algorithms such as k-means are the basic methods for community detection. However, the detection results are often not accurate enough when dealing with high-dimensional matrix when using these classical methods. In this study, a community detection algorithm based on deep sparse autoencoder (CoDDA) is proposed to improve the accuracy of community detection using high-dimensional adjacent matrix with the classical methods. First, a hop-based operation for sparse adjacent matrix is provided to obtain the similarity matrix, which can express not only the relations between nodes that are linked but also the relations between nodes that are not linked. Then, a deep sparse autoencoder based on unsupervised deep learning methods is designed to extract the features of similarity matrix and obtain the low-dimensional feature matrix which can represent the features of network topology better than similarity matrix. Finally, k-means is used to identify the communities according to the feature matrix. Experimental results show that CoDDA can obtain more accurate communities than the six baseline methods. Besides, the parameter analysis indicates that CoDDA can result in more accurate communities than the k-means algorithm which finds the communities according to the high-dimensional matrix directly.