Aiming at the uncertainty value of weight coefficient and the big error of characteristic quantification in soft assignment of characteristics quantification in Vector of Locally Aggregated Descriptor (VLAD) model, an efficient weight coefficient soft quantization assignment algorithm based on minimized reconstruction error was proposed. The sparse coding coefficients with the minimized reconstruction errors were taken as the weighting values of soft quantization assignment based on VLAD by taking the minimized reconstruction error as the standard. The image retrieval test results of database show that, compared with the mainstream VLAD feature coding algorithms, the image retrieval accuracy of the proposed algorithm can be improved about 10%, and the proposed algorithm can obtain a smaller feature reconstruction error.