A lightweight detection method for common peach tree diseases based on the FHLE-RTDETR model

• Lightweight detection method for peach tree diseases was proposed. • A dataset of multi-scale peach tree diseases in natural environments was constructed. • Partial convolution and HiLo attention enhanced the ability to capture disease features. • Esilm-neck fusion structure enhanced the model's adaptability to multi-resolution features adaptability. • The optimized RT-DETR model achieved a reduction in the number of parameters by 26%, computational load (GFLOPs) by 25.9%, and model size by 25.4%. To improve multi-scale disease detection in orchard environments with complex backgrounds while reducing model complexity, this study proposes FHLE-RTDETR, a lightweight detection model based on RT-DETR. Firstly, a dataset of common peach tree diseases was self-constructed, comprising 6,692 images of diseases, including gummosis, wood rot, shot hole disease, anthracnose, and red leaf disease. Through data augmentation, the dataset was expanded to 16,060 images to ensure the robustness of the model. Secondly, the FPCBlock module was designed to replace BasicBlock, utilizing partial convolution to reduce computational redundancy. Additionally, the high-frequency and low-frequency (HiLo) attention mechanism was integrated into the encoder to enhance detail and global feature extraction, improving the model's ability to capture key disease characteristics. Finally, the Esilm-neck feature fusion structure was introduced, combining global heterogeneous kernel selection with multi-scale convolution modules. Efficient upsampling convolution block (EUCB) and fast normalised fusion (FastNF) modules were employed to achieve efficient feature enhancement, reducing computational overhead while further improving detection accuracy. Experimental results demonstrate that FHLE-RTDETR achieved a mean average precision at 50% IoU (mAP50) of 92.1%, with precision, recall, and mAP50 increasing by 1.9%, 2.8%, and 1.7%, respectively, outperforming the original RT-DETR. Meanwhile, the model's parameters, giga floating point operations per second (GFLOPs), and size were reduced by 26%, 25.9%, and 25.4%, respectively, with an inference time of only 3.2 milliseconds per image. The proposed model surpasses baseline RT-DETR and other mainstream methods in precision, speed, and computational efficiency, and is capable of detecting diseases under different scales and complex backgrounds.