Nondestructive detection of chilled mutton freshness based on multi-label information fusion and adaptive BP neural network

Nondestructive detection of mutton freshness based on a single indicator has limitations and poor applicability. Hyperspectral imaging technology can detect a variety of information in the process of mutton freshness changes, but the establishments of spectral feature extraction and evaluation model have greater impacts on the final evaluation results. To improve the applicability and accuracy of nondestructive detection of mutton freshness, this paper proposes a multi-indicator correlation maximization spectral feature extraction algorithm and self-adaptive BP neural network classification model to detect mutton freshness nondestructively. In the experiments, 400–1000 nm hyperspectral images are collected from 140 mutton samples, and the standard values of total volatile basic nitrogen (TVB-N) and total aerobic plate count (TAC) are determined by laboratory methods. The representative spectra of mutton samples are extracted in the region of interests (ROIs). The spectral feature information is extracted using the feature of multi-indicator correlation maximization, which is proposed in this paper. The samples of calibration set and the prediction set are divided at a ratio of 3:1, and the improved three-layer BP neural network is used to evaluate the freshness of mutton. The results show that the final overall classification accuracy (OA) is 0.9378, the Kappa coefficient is 0.9096, and the AUC values for the three freshness classification results are 0.8192, 0.6766 and 0.7483, respectively. The root means square error (RMSE) of the evaluation results is 0.279. The research shows that the method proposed in this paper can achieve more accurate nondestructive detection of mutton freshness. Moreover, the research provides a new method to obtain spectral information of multiple detection indicators with hyperspectral imaging technology, which can improve the applicability and robustness of the evaluation model of single indicator.