Abstract: To address the problem that it is difficult to obtain effective deep learning classification models in the absence of a large amount of consistent distribution and labeled data in the target application scenario, a domain adaptation framework was proposed based on explicit feature alignment with domain adversarial and implicit category alignment. The advantages of the domain distribution difference approach and the adversarial learning approach were combined in this framework. For the explicit feature alignment module, considering the optimization challenges caused by large differences in domain knowledge, an incremental collaborative optimization strategy was used. The goal of this strategy is to improve the domain adaptation performance by reducing the domain differences between different semantic layers. For the implicit category alignment module, to enhance the discriminability of the target features, a self-training method was adopted to obtain pseudo-labels and overcome the problem of label noises. Besides, the accuracy of the pseudo-labels was optimized by learning the confusion matrix, thus a target domain loss function was automatically constructed. This function is able to reduce the differences between domains, while improving the accurate alignment of the feature distribution of the same categories in both the source and target domains. Experiments based on 6 cross-domain classification tasks on the Office-31 dataset and 12 sets of cross-domain classification tasks on the Office-Home dataset were conducted. Results show that the average classification accuracy of the proposed method on the transfer learning tasks improves by 11.9% and 19.9%, respectively, compared with the benchmark method, demonstrating the effectiveness of the network in domain adaptation tasks.