Hyperspectral Imaging Technology Empowers Rapid and Non-Destructive Testing of Pork Freshness

During the processing, distribution, and circulation of pork and its products, freshness is a vital indicator for measuring its quality and safety. Although traditional detection methods, such as the determination of total volatile basic nitrogen (TVB-N) and total viable count (TVC), deliver reliable results, they are cumbersome, time-consuming, and destructive to samples. Consequently, it is difficult for them to meet the demands of the modern food industry for online, rapid, and non-destructive testing.

In recent years, hyperspectral imaging technology has demonstrated significant application potential in the field of food quality assessment due to its rich information, non-contact nature, and rapid analysis capabilities. The FS-IQ-VISNIR hyperspectral camera (400–1000 nm) from CHNSpec has provided data acquisition support for this type of research.

Experimental Design and Data Acquisition

In a study published in the Journal of Food Composition and Analysis, a research team from Zhengzhou Light Industry University utilized CHNSpec's FS-IQ-VISNIR hyperspectral camera to collect visible-near-infrared hyperspectral data of pork tenderloin stored under cold refrigeration at 4°C within 14 days. A total of 112 samples were involved, covering 7 time points with 16 samples per point. The camera utilizes a push-broom imaging method, featuring a wavelength range of 400–1000 nm, containing 1200 bands, a spectral sampling interval of approximately 0.5 nm, and a spatial resolution of 1920×1920 pixels.

To enhance signal quality, the research team designed an unsupervised image preprocessing method based on spectral differences. Combined with Otsu adaptive threshold segmentation and morphological operations, this method effectively extracted the region of interest (ROI) and reduced background interference.

Dual-Branch Feature Extraction Network and Machine Learning Fusion Modeling

The study proposed a dual-branch hyperspectral feature extraction network named HFE (Hybrid-FeatureExtractor). This network consists of two parallel feature extraction channels:

• Spectral Branch: It introduces the Squeeze-and-Excitation (SE) attention mechanism to adaptively learn the weights of each band, and combines a multilayer perceptron (MLP) to extract key spectral features.
• Spatial Branch: It adopts a two-dimensional convolutional neural network (CNN) combined with residual modules (BasicBlock) and Atrous Spatial Pyramid Pooling (ASPP) modules to extract multi-scale spatial features.

The two types of features are integrated through a gated fusion mechanism and subsequently input into Partial Least Squares Regression (PLSR) and Support Vector Regression (SVR) models, respectively, to complete the quantitative prediction of TVB-N and TVC content.

Prediction Performance and Analysis

The experimental results indicate that when the HFE module is combined with PLSR and SVR, it exhibits high accuracy in predicting TVB-N and TVC content:

• HFE + PLSR: The R² for TVB-N prediction is 0.9786 with an RMSE of 2.4685; the R²  for TVC prediction is 0.9529 with an RMSE of 0.3223.
• HFE + SVR: TheR² for TVC prediction is 0.9597 with an RMSE of 0.3066.
• Compared with traditional chemometric methods (such as SG+SPA, SNV+CARS), the prediction accuracy and model stability of this method are both improved. In terms of residual prediction deviation (RPD), TVB-N and TVC reached 7.1204 and 5.1831 respectively, indicating that the model possesses strong predictive capabilities.

Model Interpretability and Key Band Identification

Through the SE attention mechanism, the research team conducted a visualization analysis of the weights of different bands in the spectral branch. The results showed that the model assigned higher weights within the 600–920 nm range, a region closely related to the optical response of protein oxidation and microbial metabolites (such as amines, aldehydes, ketones, etc.). As the storage time extended, the TVB-N and TVC content increased, and the range of these characteristic bands also expanded, indicating that the model can capture subtle spectral changes related to freshness variations.

Application Prospects

This study demonstrates that the FS-IQ-VISNIR hyperspectral camera, combined with the dual-branch feature extraction network and machine learning regression models, can achieve effective prediction of pork freshness indicators without causing damage to the samples. This method holds practical reference value for online, non-destructive testing in food processing, cold chain transportation, and retail links.

CHNSpec will continue to provide hyperspectral imaging equipment and technical support for the field of food quality and safety, helping the industry transition to more efficient and intelligent testing methods.

Product Recommendation: FS-IQ-VISNIR Portable Hyperspectral Camera

• Spectral Range: 400–1000 nm
• Spectral Resolution: 2.5 nm
• Image Resolution: 1920*1920
• Number of Spectral Channels: 1200