Citrus Disease Image Generation and Classification Based on Improved FastGAN and EfficientNet-B5

The rapid and accurate identification of citrus leaf diseases is crucial for the sustainable development industry. Because disease samples are small, unevenly distributed, difficult to collect, we redesigned generator structure FastGAN added small batch standard deviations discriminator produce an enhanced model called FastGAN2, which was used generating nutritional deficiency (zinc magnesium deficiency) images. performance existing degrades significantly when training test data exhibit large differences in appearance or originate from different regions. To solve this problem, propose EfficientNet-B5 network incorporating adaptive angular margin (Arcface) loss with adversarial weight perturbation mechanism, call it EfficientNet-B5-pro. FastGAN2 can be trained using only 50 Fréchet Inception Distance (FID) Kernel (KID) improved by 31.8% 59.86%, respectively, compared original network; 8000 images were generated (2000 black star disease, 2000 canker healthy, deficiency). Only as set train ten classification networks. Real images, not network, set. average accuracy rates networks exceeded 93%. accuracy, precision, recall, F1 scores achieved EfficientNet-B5-pro 97.04%, 97.32%, 96.96%, 97.09%, they 2.26%, 1.19%, 1.98%, 1.86% higher than those EfficientNet-B5, respectively. successfully has good generalization robustness. method study effective tool image a number samples.