Potential of deep learning and snapshot hyperspectral imaging for classification of species in meat. (November 2020)

• Record Type: Journal Article
• Authors: Al-Sarayreh, Mahmoud; Reis, Marlon M.; Yan, Wei Qi; Klette, Reinhard
• Is Part Of: Food control. Volume 117(2020)
• Publication Date: 2020-11
• Subjects: Hyperspectral -- Snapshot HSI -- 3D-CNN -- Meat classification -- Meat authenticity -- Pixel-wise classification; Food -- Quality -- Periodicals; Food -- Analysis -- Periodicals; Food handling -- Periodicals; Food industry and trade -- Quality control -- Periodicals; Aliments -- Industrie et commerce -- Qualité -- Contrôle -- Périodiques; Aliments -- Qualité -- Périodiques; Aliments -- Analyse -- Périodiques; Hygiène alimentaire -- Périodiques; Food -- Analysis; Food handling; Food -- Quality; Periodicals; Electronic journals; 664.07
• DOI: 10.1016/j.foodcont.2020.107332
• Languages: English
• ISSNs: 0956-7135

Abstract

Abstract: In the meat industry, ensuring product authenticity is important due to its impact on human diet, processing chain as well as for ensuring fair trade practices. This study aims at investigating the potential of snapshot hyperspectral imaging (HSI) and deep learning approach for red-meat classification by combining the spectral and spatial features of HSI data of red-meat products. Moreover, this study provides a comprehensive comparison between snapshot HSI and a standard line-scanning HSI system. A novel deep 3D convolution neural network (3D-CNN) model is proposed for extracting the combined features and then classifying the meat in the HSI image. An innovative graph-based post-processing method is also proposed for enhancing the prediction of the 3D-CNN approach. Results show that the 3D-CNN approach significantly enhanced the overall accuracy of state-of-the-art models. Despite limitations in the spectral information from snapshot HSI, the 3D-CNN model shows robustness in classifying red-meat with an overall accuracy of 96.9% and 97.1% for Near-Infrared (NIR) and visible (VIS) snapshot HSI, respectively. A comparison between the HSI systems revealed that state-of-the-art models are insufficient for achieving accurate classification with snapshot HSI data while the 3D-CNN achieves excellent classification accuracy on all systems by utilizing the whole image information. The current study opens the door for further research towards real-time red-meat authenticity and mobile HSI systems using snapshot HSI and deep learning models. Highlights: Snapshot HSI systems collect images at video rate and completely portable systems. Deep learning provides a deep understanding of spectral-spatial features of meat. 3D-CNN networks capture the interaction between the chemical and textural of meat. Deep 3D-CNN provides a stable accuracy on line-scanning and snapshot HSI images. Machine learning models were highly affected by the spectral range and resolution.