By designing and optimizing an objective function, and combining the distribution of samples, hash learning learns the hash codes of samples. In the existing hashing models, linear model is widely used due to its conciseness and high efficiency. For the parameter optimization of linear hashing model, a model parameter re-optimization method is propose based on similarity drive, which can improve the precision of the existing linear model-based hashing algorithms. Given a hashing method, this method is firstly run for several times with obtaining several hash matrices. Then, some bits are selected for these hash matrices to obtain a new final hash matrix based on the similarity preserving degree and a fusion strategy. Finally, this new hash matrix is used to re-optimize the model parameters, and a better hash model is obtained for out-of-sample extension. Extensive experiments are performed based on three benchmark datasets and the results demonstrate the superior performance of the proposed framework.