Developing universal classification models for the detection of early decayed citrus by structured-illumination reflectance imaging coupling with deep learning methods

Early detection of decay caused by fungal infection in citrus fruit is a major challenge for the citrus industry, as the decayed area is almost invisible on the surface of fruit. This study constructed a new detection system for structural illumination imaging combined with light-emitting diode (LED) lamp and a monochrome camera. The direct component (DC) and alternating component (AC) images were recovered by demodulating three phaseshifting pattern images under the spatial frequency of 0.25 cycles mm(--1). Compared with the DC image, the decayed area can be clearly displayed in the AC image and ratio image (i.e. AC/DC). For independent models, the classification accuracy of the decayed oranges and sugar mandarins reached 92.5% and 95.0% by combining RT images with convolutional neural network (CNN) method, respectively. However, it is time-consuming and labor-intensive to construct different models to predict the corresponding citrus variety. Thus, this study also explored the feasibility of establishing the universal classification model suitable for various citrus fruit. The classification performance of partial least square discriminant analysis and CNN models was evaluated and compared. Among all universal models, the CNN model exhibited superior performance with classification accuracies of 95.0% for independent test set including two varieties of citrus fruit (orange and sugar mandarin). For four types of citrus (orange, sugar mandarin, dekopon and Nanfeng sweet mandarin), the overall classification accuracy of the universal model was 90.6%. This study demonstrated that different varieties of early decayed citrus can be effectively identified by constructing a universal CNN model combined with structured-illumination reflectance imaging technology.