The discolorations or abnormal colors in meat and meat products during processing and storage have negative effects on their commercial value. In this study, myoglobin content (MetMb and OxyMb) in Tan mutton was rapidly detected using near-infrared hyperspectral imaging (NIR-HSI) system (900–1700 nm), and built predictive models with full wavebands (FW) based on partial least squares regression (PLSR), least-squares support vector machines (LSSVM), and back propagation neuron network (BP). To reduce the computational complexity of calibration models, feature bands were obtained by bootstrapping soft shrinkage (BOSS), variable combination population analysis coupled with iteratively retains informative variables (VCPA-IRIV), and competitive adaptive reweighted sampling (CARS), respectively. The optimized BP model based on feature wavebands with BOSS method selection displayed the best capability for predicting MetMb level (R2C = 0.8340, R2P = 0.8253 RMSEC = 3.1592, RMSEP = 3.2918). In addition, the simplified VCPA-IRIV-BP model was significant in predicting OxyMb content with R2C, R2P, RMSEC, and RMSEP values of 0.8024, 0.8680, 3.4676, and 2.7605, respectively. Results provided a theoretical reference for rapid evaluation of myoglobin content in other animal products via NIR-HSI.