Frozen-thawed pork was widely used as a raw material for processing boneless meat products. The hazardous bone fragment (1~2.5 cm) embedded in the pork can risk processing equipment and consumer health. Therefore, it is necessary to study the feasibility of multispectral imaging technology (405~970 nm) for rapid and non-destructive identification of the bone fragments embedded in frozen-thawed pork. In this work, 195 lean pork slices (LPSs) were prepared into 65 samples of boneless LPSs, 65 samples of bone fragments embedded in the surface of LPSs and 65 samples of bone fragments embedded in the inner of LPSs, and then the multispectral images of them were captured after freeze-thaw treatment. These images were segmented by canonical discriminant analysis (CDA) and converted into two types of regions of interest(ROIs-1 and ROIs-2), then extracted their spectral and image information. Finally, the identification models of hazardous bone fragments embedded in the frozen-thawed LPSs were established by support vector machine (SVM) and neural network (NN). The results showed that the whole spectra extracted from ROIs-2 had better identification ability of bone fragments than that extracted from ROIs-1 and could be used to establish SVM and NN models with 100% accuracy, indicating that the region segmentation was closely related to model accuracy. The bone fragments in pork could be identified with 100% accuracy using the spectra extracted from ROIs-2 at six key wavelengths (505, 590, 700, 850, 890 and 970 nm) that were selected by successive projection algorithm (SPA), implying that the testing efficiency was further improved. The image information had a significant advantage because it could establish the SVM model with 93.8% accuracy and the NN model with 93.33% accuracy for identifying the bone fragments that were lower than those established by the spectral information and obtain visible results. In conclusion, the bone fragments embedded in the frozen-thawed pork could be precisely identified based on multispectral imaging technology, which would provide a theoretical basis for industrial online detection.