Although dictionary learning mostly uses linear functions to capture potential features of data, this method cannot fully extract the inherent feature structure of data. Deep learning has received widespread attention in recent years due to its outstanding feature representation ability. Therefore, this study proposes a nonlinear feature representation strategy combining deep learning with dictionary learning, i.e., deep neural network-based dictionary learning (DNNDL). DNNDL integrates the dictionary learning module into the traditional deep learning network structure and simultaneously learns the data dictionary and the sparse representation coefficients on it in the low-dimensional embedded space mapped by the autoencoder, thereby achieving end-to-end potential data feature extraction. It can generate compact and discriminant representations of existing data as well as out-of-sample point data. DNNDL not only is a new deep learning network structure but also can be regarded as a unified framework of dictionary learning and deep learning. A large number of experiments on four real data sets show that the proposed method has a better data representation capability than those of conventional methods.