Abstract: Taking the whole chicken breast as the research object,an online near-infrared spectroscopy(NIR)system was used to collect the spectra in the wavelength range of 900~1650 nm to explore the quantitative relationship between the spectral information and the total viable count(TVC). The collected original spectral information was preprocessed using five pretreatments including Gaussian filter smoothing(GFS)correction to construct partial least squares(PLS)models. The regression coefficients(RC)and successive projections algorithm(SPA),respectively,were used to select optimal wavelengths to build optimized PLS model and multiple linear regression(MLR)model. The results showed that the GFS-PLS model based on GFS spectra performed best in predicting TVC with r_P of 0.964 and RMSE_P of 0.806 lg CFU/g. SPA-GFS-MLR model based on the 25 optimal wavelengths(907.0,913.7,923.8,927.2,937.2,937.2,947.3,974.0,987.3,997.3,1007.3,1040.4,1080.1,1099.9,1132.9,1155.9,1185.5,1215.0,1241.2,1270.6,1358.2,1380.8,1403.3,1419.3,1578.9 and 1615.2nm)selected by SPA had similar accuracy and stability,with r_P of 0.944 and RMSE_P of 1.022 lg CFU/g,compared with GFS-PLS model. It is demonstrated that the on-line NIR system can be used to detect the TVC in the whole chicken breast in a rapid way.