Color is the first perceptual indicator for evaluating frozen meat quality， and establishing an in-situ method for evaluating the color of frozen pork will be critical to stabilize the consumer market. This work proposed a real-time method for detecting the color of frozen pork based on the image threshold segmentation algorithms including canonical discriminant analysis (CDA) and global threshold segmentation (GTS). A total of 120 multispectral images of frozen pork were segmented. These segmented images were then converted into corresponding spectra， which were applied to establish calibration models for predicting the color characteristics (L， a and b) of frozen pork by using various algorithms such as successive projections algorithm (SPA) and partial least squares regression (PLSR). The results showed that three optimal models for predicting L， a and b values were built by combining the GTS with SPA and PLSR， and their corresponding determination coefficients (R2) and residual prediction deviations (RPD) were 0.9565 and 4.7745， 0.9593 and 4.6265， and 0.9570 and 4.2126， respectively. Results of high accuracy and robustness would provide a theoretical basis for non-destructive and rapid detection of the color of frozen pork in industrial practice.