Detecting blueberry leaf diseases early is essential for safeguarding crop quality and yield. Traditional methods for this task often involve feature preprocessing combined with machine learning techniques such as K-NN, decision trees, or support vector machines. Recently, there has been a growing interest in deep learning approaches to enhance the performance of existing leaf disease detection systems. While convolutional deep learning methods have shown promising results under standard evaluation conditions, their performance tends to degrade in the presence of noise, such as Gaussian noise. Given the successful application of transformers in the vision domain, this study aims to investigate the use of transformers combined with a novel local descriptor to improve plant leaf disease detection performance under challenging conditions. Experimental results demonstrate that the proposed system achieves stable and superior performance even in noisy environments, outperforming existing methods, including YOLOv5 and YOLOv8, by 2.9% and 4.4% in terms of accuracy, respectively.