Abstract: In order to effectively solve the training and deployment implementation of large-scale classifiers involved in deep neural network to solve direction of arrival (DOA)estimation, we propose to decompose the traditional one-hot classifier into small co-prime classifiers with relatively small label, and then reconstruct the original one-hot label by combining the classification results of co-prime classifiers. First, the original label is decomposed into a number of small co-prime labels. The corresponding category of the small labels is the result of the remainder of the prime number of the original label. Secondly, by training each co-prime classifier independently and in parallel, the difficulty of training classifiers under the condition of large categories is reduced. The simulation results show that compared with the one-hot classifier, the co-prime classifiers has lower complexity and is easy to train. In addition, DOA estimation using co-prime classifiers can achieve super-resolution and the estimation accuracy is higher than that of one-hot classifier and sparse Bayesian learning method.