LGM-Net: Wheat Pest and Disease Detection Network Based on Local Global Information Interaction and Multi-Level Feature Fusion

Abstract

Wheat, as one of the most critical and widely consumed crops globally, is of irreplaceable importance to the food supply and agricultural economy. However, disease problems in wheat leaves often have a significant negative impact on the growth and yield of the crop. Therefore, this paper proposes a wheat pest and disease detection network based on local global information interaction and multi-level feature fusion, aiming to improve the accuracy and efficiency of wheat pest and disease detection. Firstly, this paper designs a lightweight feature interactive network (LFI-Net) to fully extract the local and global features of wheat leaf diseases and improve the detection of wheat pest and disease targets under noise interference. Secondly, this paper proposes a Multi-level Path Aggregation Network (MPA-Net), which uses the features output from all levels of the backbone network to construct a four-level node network that reduces layer by layer to improve the identification of multi-scale features in wheat pest and disease targets. Finally, this paper designs a decoupled task-adaptive detection head, including a task-aware layer dominated by channel attention and a feature learning layer dominated by deformable convolution. This model significantly improves the detection efficiency and accuracy of wheat diseases and insect pests, so as to take timely control measures and reduce the yield loss caused by diseases and insect pests, which is of great significance for ensuring global food security, and also provides strong technical support for the development of intelligent agriculture and precision agriculture. The experimental results showed that 96.5% and 98.5% mAP@.5 were obtained on the FDWD and SDWD datasets, respectively, and an FPS of 51 was achieved, which is better than other mainstream detectors.