Abstract: To evaluate the effectiveness of a deep learning which is based intelligent assisted hyperspectral imaging system on the detection of pork freshness indicators, volatile basic nitrogen (TVB-N), total viable count (TVC), and 900~2500 nm near-infrared spectral data were collected from pork which were refrigerated at 4 ℃ for 12 days. Based on Python's TensorFlow and Keras platform, hyperspectral data was processed and a quantitative detection model of deep neural network was also established. And the characteristic spectral bands related to pork freshness were selected by genetic algorithm (GA). The results showed that the performance of the spectral model could be improved significantly by selecting the band of genetic algorithm. When the number of spectral bands reached 35 and 50, the prediction accuracy of GA+ANN model was higher than that of full-band linear regression model. The predictive performance of TVC was better than that of TVB-N, and the best R_p² and RMSEP of TVC were 0.877 and 0.575, respectively. The best R_p² and RMSEP for TVB-N were 0.826 and 1.01, respectively. In addition, it was also found that the NIR band selected by genetic algorithm had a high coincidence with the molecular vibration absorption bands of meat, such as O-H, N-H, C=O and so on. This study provides a new method which can be used for processing the near-infrared and hyperspectral data, and also provides a technical reference for rapid nondestructive testing of pork and other meat freshness.