Review of Using Various Deep Learning Techniques and Cycle-GANs with Transformer for Disease Detection and Classification in Plant Leaves

In the upcoming years, agricultural yields should rise by 50–60 % to assure food security and feed the world's rapidly expanding population. Considering the importance of agriculture, it is imperative to look at methods that can enhance planting practices, track agricultural conditions, identify plant illnesses, and streamline harvesting procedures. The significance of these fundamental elements has prompted in-depth investigation and the creation of numerous studies that offer solutions to these problems. To improve productivity, identifying and preventing plant diseases is crucial. For increased agricultural yields, cost savings, and decreased crop losses, early disease diagnosis is essential. Deep learning and other computer vision-based technologies can help diagnose diseases early. However, the commonality of illness traits in images and interference from the external environment pose hurdles to existing machine vision approaches. Most of the previous work on image classification used plant leaf images captured in controlled environment, having similar background, which produced good results on the same data-set but failed to generalise on the images captured in real conditions. To solve this problem real background images ere used to train the classifiers which increased the generalisation significantly on same data-set but still performed poor on cross data-set generalisation. Attempts are done to solve the problem of data having no background diversity (Background, lighting, camera setting, distance and angle of capture), by data augmentation using GANs, Specifically by using unpaired image to image translation to translate healthy images into diseased images. Earlier models were based on CycleGAN and could not increase the accuracy significantly as CycleGAN translated the latent features also with the disease region. No study within our knowledge has employed Swin transformers in the generator network to augment the CycleGAN model's ability to capture the distribution of the diseased dataset. Furthermore, the use of a transformer-based Image-to-Image translation model reveals issues related to the generation of artifacts, particularly in areas with excessively high or low pixel values. This phenomenon was noted in the present study when applying existing models to plant leaf data augmentation. Thus, by evaluating good fid, psnr, and ssim values using the Swin transformer cycle-GAN, good quality images are obtained and the background diversity issue is resolved. Therefore, classification accuracy is increased by using both actual and produced images to aid in the process. To offer farmers effective solutions for safeguarding their crops and future applications, this paper assesses various methods, delineating their advantages and disadvantages. It proposes the utilization of a hybrid deep learning architecture along with diverse strategies for the detection and classification of plant leaf diseases.