Practical cucumber leaf disease recognition using improved Swin Transformer and small sample size

Abstract

The deep learning methods based on convolutional neural network (CNN) have been widely explored in dataset augmentation and recognition of plant leaf diseases. The recently developed transformer-based models such as Swin Transformer (SwinT) show competitive and even better performance on various visual benchmarks compared with CNN due to their inherent attention mechanism and ability to learn long-range dependency among pixels. In this paper, a backbone network based on improved SwinT was proposed and applied to the data augmentation and recognition of practical cucumber leaf diseases. Firstly, the patch partition of SwinT was improved by step-wise small patch embeddings for enhancing the ability of feature extraction without increasing the number of parameters. Secondly, the leaf extraction module composed of the proposed backbone network and Grad-CAM was integrated into the Generation Adversarial Network (GAN) to construct STA-GAN (SwinT-based and Attention-guided GAN), which generated diseased spots only in the leaf region of healthy images with complex background for augmenting the disease dataset. Finally, by means of transfer learning, the augmented datasets were used to train the recognition model of cucumber leaf diseases with the proposed backbone network. From the experimental results, it has been demonstrated that STA-GAN exhibited stronger ability to generate high-quality images than LeafGAN, even only approximated when LeafGAN consumed much more training images. Additionally, with STA-GAN, the disease recognition accuracies reached 98.97%, 96.81%, 94.85% and 90.01% when improved SwinT, original SwinT, EfficientNet-B5 and ResNet-101 were employed as the backbone of recognition model respectively, increasing by 2.17%, 3.62%, 2.13% and 11.23% compared with LeafGAN, revealing that the approaches based on improved SwinT could indeed help in boosting the performance of both data augmentation and recognition of practical cucumber leaf diseases. The proposed approach has the potential of dealing with the common challenge of insufficient data size and complex background in other similar plant science tasks.