Image Generation of Tomato Leaf Disease Identification Based on Adversarial-VAE
Yang Wu, Lihong Xu
Abstract
The deep neural network-based method requires a lot of data for training. Aiming at the problem of a lack of training images in tomato leaf disease identification, an Adversarial-VAE network model for generating images of 10 tomato leaf diseases is proposed, which is used to expand the training set for training an identification model. First, an Adversarial-VAE model is designed to generate tomato leaf disease images. Then, a multi-scale residual learning module is used to replace single-size convolution kernels to enrich extracted features, and a dense connection strategy is integrated into the Adversarial-VAE networks to further enhance the image generation ability. The training set is expanded by the proposed model, which generates the same number of images by training 10,892 images of 10 leaves. The generated images are superior to those of InfoGAN, WAE, VAE, and VAE-GAN measured by the Frechet Inception Distance (FID). The experimental results show that using the extension dataset that is generated by the Adversarial-VAE model to train the Resnet identification model could improve the accuracy of identification effectively. The model proposed in this paper could generate enough images of tomato leaf diseases and provide a feasible solution for data expansion of tomato leaf disease images.