FWDGAN-based data augmentation for tomato leaf disease identification

There are few publicly available large-scale datasets for tomato leaf disease detection using deep learning, the workload associated with manual data collection is high, and conventional supervised data augmentation methods (such as image rotation, flipping, and shifting) are challenging to attain satisfactory results with. While Deep Convolutional Generative Adversarial Networks (DCGAN) is a popular unsupervised data augmentation method based on deep learning, the data augmentation method based on DCGAN may have several drawbacks, including poor image quality and excessive computing resource consumption. Given this, a method based on Fast WDBlock based GAN (FWDGAN) was proposed in this paper. For the network's generator, a wide and deep feature extraction block (WDBlock) with a two-path strategy was designed, combining the extracted depth feature based on ResNet and the extracted global feature based on InceptionV1. By incorporating WDBlock into the generator, the quality of the generated tomato leaf disease images was improved. For the network's discriminator, the Depthwise separable convolution Discriminator (DSC-Discriminator) that significantly reduced the model's parameters without impairing the network's performance was constructed. Finally, the SeLU activation function was used selectively to improve the training stability of the network. Comparative experiments demonstrated that FWDGAN could generate higher-quality data, with FID scores of 193.998, 264.704, 260.594, and 161.436 for healthy tomato leaf image, Leaf Mold, Septoria leaf spot, and Yellow Leaf Curl Virus generated by FWDGAN, respectively. Furthermore, the total number of parameters in FWDGAN was approximately one-third less than that of DCGAN.

Automated summary

This paper proposes a FWDGAN method based on WDBlock for generating high-quality tomato leaf disease images. By combining ResNet and InceptionV1 for feature extraction, and using DSC-Discriminator, FWDGAN outperforms DCGAN in terms of data quality and parameter quantity.