GrapeGAN: Unsupervised image enhancement for improved grape leaf disease recognition

Grape leaf disease seriously affects the yield and quality of grapes. Limited by actual conditions, collecting a large number of grape disease images is time-consuming and labor intensive, which makes it difficult to train grape disease identification models with excellent performance. Currently, using generative adversarial networks (GANs) to generate grape leaf images is a popular method. Unfortunately, the leaf disease images generated by conventional GANs are not clear enough and the structural integrity is insufficient. To address this problem, a novel architecture named GrapeGAN is proposed in this paper. First, suppress the loss of texture detail information during image generation, a U-Net-like generator is designed by integrating convolutions with residual blocks and reorganization (reorg) methods. Simultaneously, the concatenation (concat) method is used in the generator to retain more scale texture information. Then, to make the generated grape images structurally complete and avoid petiole and leaf structure misalignment, a discriminator is designed with a convolution block and capsule structure. Convolution is used to extract general features, and the capsule structure encodes the spatial information and the probability of the presence of spots. In subsequent experiments on the same raw data, GrapeGAN is compared to WGAN and DCGAN, and the results show that GrapeGAN outperforms the comparative models. Specifically, the Fre & PRIME;chet inception distance (FID) is 5.495, and the neural image assessment (NIMA) is 4.937 +/- 1.515. Moreover, four convolutional neural network (CNN) recognition models are used to identify the generated grape leaf diseases. The results demonstrate that the recognition accuracy of grape leaf disease images generated by the GrapeGAN is higher than 86.36%, and the identification accuracy of VGG16 and InceptionV1 achieve 96.13%. In summary, the experimental results show the effectiveness of GrapeGAN, which proves that GrapeGAN can efficiently detect grape leaf disease detection.