The task of unsupervised person re-identification (Re-ID) is to transfer the knowledge learned in the source domain with no labels to the target domain with no labels. Due to the significant differences in the background of different datasets, the trained model is challenging to extract person features accurately on unsupervised domain adaptive (UDA). Most UDA methods for person Re-ID use single-image representation (SIR) during the feature extraction. These methods might ignore the difference among the cross-view images with the same identity. For this problem, the joint learning image representation strategy for unsupervised domain adaptation (JLIRS-UDA) is proposed, which takes cross-image representation (CIR) into account for UDA. The network architecture of JLIRS-UDA consists of two networks with branching networks. Each network consists of a shared network and two branching networks, the SIR branch and CIR branch. The two branching networks aim to learn the SIR and CIR, respectively. To ensure the accuracy of the pseudo-label generation, the segmenting dynamic clustering (SDC) method is proposed, which divides the training process into two phases. Precisely, in the first phases, SDC adopts the single image features in the clustering phase to ensure that accurate feature details can be learned. In the second phase, SDC fuses SIR and CIR as the final feature for clustering to interactively promote the SIR branch and CIR branch. JLIRS-UDA learns the SIR and CIR jointly in the UDA task training phase. Compared with the state-of-the-arts methods, the strategy proposed in this paper has achieved a significant improvement of 7.1% mean average precision (mAP) on the tasks of Market-1501(Market) to DukeMTMC- Re-ID (Duke). The slightest improvement in accuracy also achieved +0.8% on Market to MSMT17 (MSMT).