Hyperspectral Imaging Combined with Convolutional Neural Network for Rapid and Accurate Evaluation of Tilapia Fillet Freshness

Abstract

The purpose of this work is to achieve rapid and nondestructive determination of tilapia fillets storage time associated with its freshness. Here, we investigated the potential of hyperspectral imaging (HSI) combined with a convolutional neural network (CNN) in the visible and near-infrared region (vis-NIR or VNIR, 397−1003 nm) and the shortwave near-infrared region (SWNIR or SWIR, 935−1720 nm) for determining tilapia fillets freshness. Hyperspectral images of 70 tilapia fillets stored at 4 ℃ for 0–14 d were collected. Various machine learning algorithms were employed to verify the effectiveness of CNN, including partial least-squares discriminant analysis (PLS-DA), K-nearest neighbor (KNN), support vector machine (SVM), and extreme learning machine (ELM). Their performance was compared from spectral preprocessing and feature extraction. The results showed that PLS-DA, KNN, SVM, and ELM require appropriate preprocessing methods and feature extraction to improve their accuracy, while CNN without the requirement of these complex processes achieved higher accuracy than the other algorithms. CNN achieved accuracy of 100% in the test set of VNIR, and achieved 87.30% in the test set of SWIR, indicating that VNIR HSI is more suitable for detection freshness of tilapia. Overall, HSI combined with CNN could be used to rapidly and accurately evaluating tilapia fillets freshness.