GMOPNet: A GAN-MLP two-stage network for optical properties measurement of kiwifruit and peaches with spatial frequency domain imaging

Spatial frequency domain imaging (SFDI) is an imaging technique using spatially modulated illumination for measurement of optical properties. Conventional SFDI methods require capturing at least six images, making it time-consuming. This study presents a Generative Adversarial Network-Multi-Layer Perceptron (GAN-MLP) two- stage network (GMOPNet) for extracting high-precision optical properties of kiwifruit and peaches from a single SFDI image, enabling real-time continuous wide-band SFDI. The GMOPNet we proposed leverages the GAN to predict diffuse reflectance, followed by the MLP with Monte Carlo prior knowledge to predict optical properties. Our method achieves mean absolute percentage errors (MAPE) of 5.91% for the absorption coefficient (mu a) and 5.23% for the reduced scattering coefficient (mu s'), reducing acquisition and processing time significantly, with single inference taking 31.13 ms. The MAPE of the mu a and the mu s ' were 6.73% and 6.34% for kiwifruit and 5.80% and 6.65% for peaches, respectively.