For large-scale image retrieval, hashing algorithms are one of the most widely used methods due to their computational and storage efficiency. Compared with most of the existing data-dependent pair/triplet-based hashing methods, the hashing method based on central similarity quantization can optimize the global similarity more efficiently and alleviate the problem of missing the global nature of the data distribution. However, there still exists a lack of expression of the feature capability. Because of the different objective functions, there is an incompatible conflict between the optimal clustering position and the ideal hash position, leading to serious ambiguity and erroneous hashing after binarization. Therefore, we employ a hinge embedding function to explicitly force the termination of the metric loss to prevent negative pairwise infinite discretization. In addition, the performance difference of the models used in deep hash retrieval can also limit the efficiency of retrieval. To solve this problem, we propose an integration learning framework for image retrieval, which can learn compact hash codes by hash center constraints. We introduce the integration strategy and integrate the retrieval results using the weighted average method. Comprehensive experiments on three benchmark datasets, MS COCO, VOC2012, and ImageNet, show that the present framework has superior average accuracy mean on different lengths of hash code retrieval.