Sliding-window enhanced olfactory visual images combined with deep learning to predict TVB-N content in chilled mutton

Abstract

A novel data enhancement method for olfactory visual images was proposed in this study, combined with deep learning to achieve the accurate prediction of total volatile basic nitrogen (TVB-N) content in chilled mutton. Specifically, the sliding-window was defined and used to separately extract different regions of interest from each sensing region by encoding and decoding the sliding position information, so the olfactory visual image was enhanced. This enhancement method considered the position shift and uneven colour presentation of sensitive points during the preparation and reaction of olfactory visualization sensor array. Based on the enhanced images, three advanced deep learning models (InceptionNetV3, ResNet50 and MobileNetV3) were established, and compared with three traditional machine learning models of partial least squares regression (PLSR), support vector regression (SVR) and random forest (RF) based on manually extracted colour space features. By comparison, deep learning models of InceptionNetV3, ResNet50 and MobileNetV3 had better predictive performance, and the optimal prediction results were obtained by the MobileNetV3 model. The determination coefficient (R²), root-mean-square error (RMSE) and relative prediction deviation (RPD) of the best prediction model for test set were 0.97, 2.42 mg/100 g and 5.82, respectively. The results demonstrated that the combination of olfactory visualization sensor array and the lightweight MobileNetV3 can stably and effectively predict the TVB-N content in chilled mutton, and has great potential for on-site evaluation of mutton freshness.