Abstract: The data sources in image recognition are complex, and there are differences in the distribution of data from different domains. To improve the recognition ability of cross-domain images, this paper proposes a projection reconstruction based domain-adaptive dictionary pair learning (PRDDPL) method. This method employs cross-reconstruction technique to construct new source and target domains, uses synthesis and analysis dictionary pairs to align samples in different domains, and utilizes the association of dictionary atoms with class information to transfer discriminative information from the source domain to the target domain. At the same time, the discriminative ability of the dictionary is improved by analyzing the dictionary constraints. By minimizing the linear classification error of each class of data and maximizing the difference between different classes, the discriminative ability of the sparse coefficients is improved through the classification discriminant of the source domain and target domain. The experiment is performed on the Cal tech-256 and Office datasets. The experimental results show that using Surf features the proposed method achieves the classification accuracy of 54.99% and 58.64% in single source domain and multi-source domain experiments, respectively. Using deep features, the proposed method achieves the classification accuracy of 91.43% and 92.66% in single source domain and multi-source domain experiments, respectively.