The work aims to evaluate the freshness of large yellow croaker (Pseudosciaena crocea) by near infrared spectroscopy (NIRS) during ice storage for different time. Total viable counts (TVC) were used as an index to evaluate the freshness and the partial least squares (PLS) were adopted to construct the model based on the pretreatment methods like Mean Center (MC), Normalization by Closure (Ncl), Standard Normal Variate (SNV), Derivative 1st (Db1), Derivative 2nd (Db2) and Multiplicative Scatter Correction (MSC). By comparing the correlation coefficient and standard deviation of calibration set and validation set, the best TVC model for rapidly predicting the freshness of large yellow croaker (Pseudosciaena crocea) in ice storage was established. Ncl could eliminate the spectral noise better and improve the prediction ability of the model. After pretreatment by Ncl, the best modeling effects could be achieved by PLS. The correlation coefficient of calibration set was 0.9095, the correlation coefficient of standard error for calibration (SEC) was 0.5872, the correlation coefficient of verification set was 0.8858, and the standard error of prediction (SEP) was 0.6615. The correlation coefficient of model was higher than 0.9. The result shows that the model has good prediction accuracy, and has a good application prospect in the freshness detection and quality evaluation of large yellow croaker (Pseudosciaena crocea).