The total counts of Enterobacteriaceae and Pseudomonas spp. (EPC) appeared on edible salmon flesh were determined by using near-infrared (NIR) (900–1700nm) hyperspectral imaging. Three-dimensional hyperspectral images (x,y,λ) of salmon samples were acquired at different storage time. Spectra (λ) extracted in reflectance (R) unit and two other transformed spectra units, absorbance (A) and Kubelka–Munck (KM), were prepared to relate to the measured EPC data by using partial least square (PLS) regression. Based on the three spectra parameters, three full wavelength PLS models defined as FR-PLS, FA-PLS and FKM-PLS were developed with all correlation coefficients of prediction (RP) over 0.900. To simplify these models, wavelengths holding the most important information were selected by executing competitive adaptive reweighted sampling (CARS) algorithm. Better performance was found in the resulting simplified R-PLS model (defined as FRS-PLS model) which was established with only nine important wavelengths (931, 1138, 1175, 1242, 1359, 1628, 1641, 1652 and 1655nm) selected from R spectra. The absolute difference between root mean square errors of calibration (RMSEC) and prediction (RMSEP) in the FRS-PLS model was 0.063, less than half (44%) of that of the original FR-PLS model. By applying the FRS-PLS model to the 2-D images (x,y), EPC distribution maps were generated to visualize the spatial variation of EPC and the adaptability of the FRS-PLS model for EPC evaluation was further demonstrated with these distribution maps in which different colors indicated different degrees of EPC contamination. To sum up, NIR hyperspectral imaging technology shows a great potential to predict the EPC contamination in salmon flesh. In view of the results obtained from this study, a multi-spectral imaging system could be developed and further refined for online detection applications in fish industry.