In the field of person re-identification, re-ranking as a post-processing technique is crucial for improving accuracy. Current strategies primarily rely on mutual nearest neighbor relationships as constraints. However, due to the limitations of model performance, confusing samples are easily introduced, resulting in poor re-ranking effectiveness. To address this issue, an attribute-similar K-reciprocal nearest neighbors re-ranking strategy is proposed. Based on the K-reciprocal nearest neighbors, this strategy uses the attribute similarity of same-source person images as a constraint to build a robust candidate image set, thereby enhancing the re-ranking effect. Given the strong dependence of this strategy on attribute recognition performance, an attribute relevance and part association network is proposed. Firstly, the model couples pedestrian attributes with local regions, recognizing the attributes corresponding to each part while extracting local features. Secondly, an attribute relevance module is proposed to use the inherent correlation between attributes to correct incorrectly predicted attributes. Thirdly, an attribute similarity loss is proposed, which uses image pairs from different views to associate and complement occluded attributes, further improving the accuracy of attribute recognition. Finally, the attribute recognition results are applied to the attribute-similar K-reciprocal nearest neighbors re-ranking strategy to obtain better re-ranking results. Experiments results on the Market-1501 and DukeMTMC-reID datasets demonstrate that the proposed method can significantly improve person re-identification performance.