Noninvasive Diagnosis of Seedless Fruit Using Deep Learning in Persimmon

Noninvasive diagnosis of internal traits in fruit crops is a high unmet need; however it generally requires time, costs, and special methods or facilities. Recent progress deep neural network (or learning) techniques would allow easy, but highly accurate with single RGB images, the latest applications enable visualization “the reasons for each diagnosis” by backpropagation networks. Here, we propose an application learning image on classification traits, this case seedlessness, persimmon (Diospyros kaki). We examined seedlessness using four convolutional networks (CNN) models various layer structures. With only 599 pictures ‘Fuyu’ from apex side, successfully made binary seedless seeded fruits up to 85% accuracy. Among CNN models, VGG16 model simplest structure showed highest accuracy 89%. Prediction values were significantly increased proportion seed numbers all models. Furthermore, explainable AI methods, such as Gradient-weighted Class Activation Mapping (Grad-CAM) Guided Grad-CAM, allowed parts patterns contributing diagnosis. The results indicated that finer positions surrounding apex, which correspond hypothetical bulges derived seeds, are index fruits. These suggest novel potential noninvasive simple images also provide insights into previously unrecognized features seeded/seedless