Novel feature extraction in laser light backscattering imaging for real-time monitoring of quince moisture content during hot-air drying

Abstract

Non-contact and accurate determination of product moisture content during drying is essential for maintaining quality and evaluating drying performance. In this study, a specific drying chamber was equipped with a laser light backscattering imaging (LLBI) setup to capture real-time backscattering images of quince slices. Two diode-pumped lasers, operating at green (532 nm) and near-infrared (NIR) (980 nm) wavelengths, were implemented for this purpose. In addition to extracting color features from backscattered regions, state-of-the-art shape features were also extracted from both saturated and backscattered regions of the lasers by measuring radial profiles (RPs). Furthermore, two pre-trained convolutional neural networks, namely ResNet50 and VGG19, were utilized to extract new deep features. The color and shape features of both lasers were assessed individually and in a fusion strategy to maximize the predictability of moisture content using two regression methods: partial least squares (PLS) and artificial neural networks (ANN). The results demonstrated excellent predictability of moisture content when color and shape features of the green laser were fused into an ANN model (SDR of 3.00). However, the NIR laser yielded moderate predictions individually, particularly when utilizing VGG19 deep features (SDR of 2.08). Moreover, the fusion of color and shape features from both lasers exhibited strong synergy, resulting in the best ANN predictive model (R²_p of 0.920, RMSEP of 7.24%, and SDR of 3.56). Through the utilization of these novel features, this study highlights the significant potential of the LLBI technique for real-time monitoring of moisture content in quince slices during drying.