IMLL-DETR: An intelligent model for detecting multi-scale litchi leaf diseases and pests in complex agricultural environments

Abstract

Accurately detecting and identifying litchi leaf diseases and pests is vital for achieving effective control. However, litchi leaf disease and pest detection methods have faced several challenges, such as multi-scale targets, high-density targets, and interference from weather, which increase the difficulty of detection for general network models. Therefore, this study proposes a highly accurate and intensely robust target detection model based on an improved Real-Time DEtection TRansformer (RT-DETR) called the Intelligent Multi-scale Litchi Leaf diseases and pests DETR (IMLL-DETR). To mitigate the impacts of these uncertainties in complex agricultural environments, this study proposes a Multi-scale Dynamic convolutional Gated Attention module (MDGA). The MDGA module enhances the focus on local multi-scale targets by extracting features from targets at different scales via multi-branch depth-wise strip convolutions. It also introduces Dynamic convolution (DynamicConv) to obtain the global key contextual information to prevent the model from being dominated by local features, reduce the impacts of complex backgrounds and the relationships between targets, and then dynamically adjust the local multi-scale information and global contextual key information through a gating mechanism to achieve an ideal balance between them, thereby effectively improving the accuracy and robustness of the litchi disease and pest detection process. Additionally, to overcome the loss of information about small targets and detailed features when litchi leaf disease and pest features are propagated over long distances in the network, the IMLL-DETR introduces the P2 detection head. Finally, to improve the accuracy of small target localization and the distinctions between closely spaced targets, the NWD_Focaler-GIoU localization loss function is restructured. To evaluate the performance of the IMLL-DETR in complex agricultural environments, image data concerning five common litchi leaf diseases and pests were collected from natural agriculture environments under three weather conditions, and a multi-scale target dataset named LEADA was constructed. The experimental results showed that on the LEADA, the AP₅₀, AP_50:95, and AR_50:95 of the IMLL-DETR reached 84.4 %, 49.1 %, and 56.9 %, respectively, which were 3.8 %, 2.7 %, and 2.0 % higher than those of the baseline model. For the large, medium, and small targets contained in the LEADA dataset, the AP_50:95 values of the IMLL-DETR reached 82.1 %, 67.1 %, and 41.1 %, respectively, marking improvements of 8.3 %, 2.6 %, and 3.2 % over the baseline model. In addition, the AP₅₀ values produced by the IMLL-DETR for scenarios with sunlight interference, rain interference, and high-density targets improved by 5.9 %, 3.3 %, and 4.9 %, respectively. These results demonstrate that the IMLL-DETR has high accuracy and strong robustness. Finally, the experimental results obtained on other datasets and other models validated the superiority of the IMLL-DETR. This study is expected to provide adequate technical support for controlling and managing litchi leaf diseases and pests.