A method for freshness detection of pork using two-dimensional correlation spectroscopy images combined with dual-branch deep learning

This study proposes a new visible near-infrared spectral analysis method by abandoning the traditional feature engineering-based method. It presents a dual-branch convolutional neural network (CNN) with bilinear pooling, combined with synchronous-asynchronous two-dimensional correlation spectroscopy (2D-COS) images, for quantifying pork freshness. 2D-COS images revealed spectral correlations of pork with different freshness at different bands, effectively separating overlapping bands and amplifying spectral differences. A dual-branch CNN using a bilinear pool integrates synchronous and asynchronous features to capture the interactions between features for quantitative analysis. Results demonstrate that the dual-branch CNN achieves high accuracy (R2p=0.9579 and RMSEP=0.8093 mg/100 g) for Total Volatile Basic Nitrogen (TVB-N) content prediction. The mechanism of TVB-N prediction is explained using Gradient-weighted Class Activation Mapping (Grad-CAM), demonstrating the reliability of the proposed method to replace human experience for feature extraction and modeling analysis. In conclusion, this study proposes a new approach for food inspection tasks that is convenient, efficient and human-expertise-independent.