ERL-RTDETR: A Lightweight Transformer-Based Framework for High-Accuracy Apple Disease Detection in Precision Agriculture

Abstract

Apples are deeply favored by consumers for their crisp and sweet taste and play a significant role in agricultural production. However, apples often suffer from infections by various pathogens during their growth process, severely impacting fruit quality and yield, and subsequently causing economic losses. Therefore, timely detection and accurate intervention against diseases during apple growth are crucial for improving harvest management efficiency and economic benefits. Nonetheless, current research primarily focuses on the identification of single diseases, lacking multi-disease detection capabilities. This limitation results in inadequate timeliness and accuracy in disease management, thereby restricting practical application effectiveness. Additionally, apple disease detection models need to balance high accuracy, rapid response, and lightweight design to reduce hardware costs and application thresholds. To address these challenges, this paper proposes a lightweight detection model named ERL-RTDETR, which is based on RT-DETR. First, a dataset containing 3096 images of apple-leaf diseases was constructed, encompassing different camera angles, time spans, and lighting conditions in complex environments. Subsequently, by introducing an Efficient Multi-scale Attention (EMA) mechanism and integrating it with the backbone network, we designed a new feature extraction module (BasicBlock_EMA) to enhance the capture of fine-grained features. Meanwhile, in the neck network, the traditional convolutional module was replaced with a Lightweight Adaptive Extraction module (LAE), and a Generalized Efficient Lightweight Attention Network (GELAN) was introduced to optimize the convolutional blocks, thereby improving the model's training efficiency and detection performance for subtle targets. The construction of the ERL-RTDETR model was completed while ensuring detection accuracy and reducing model complexity. Experimental results demonstrate that ERL-RTDETR achieves a balanced performance in apple disease detection tasks, with a detection precision of 94.5% on the test set (a 3.2% improvement compared to RT-DETR) and increases in mAP50 and mAP50:95 by 2.7% and 2.2%, respectively. Simultaneously, the GFLOPs were reduced by 5.9 GFLOPs (a decrease of 10.3% compared to RT-DETR). In summary, the proposed ERL-RTDETR model provides an efficient, lightweight, and accurate method for apple disease detection, serving as an important reference for research and practical applications in related fields.