• A novel MEGE-DETR is proposed for dense small-object pest detection. • MEGFNet enhances multi-scale edge features with fewer parameters. • Polarity-aware Linear Attention Encoder preserves details in complex scenes. • Learnable wavelet upsampling improves edge restoration performance. • A new dataset with five pest types is constructed for evaluation. Accurate and rapid detection of dense small-object pest is of vital importance for crop protection. However, existing methods often suffer from limited feature extraction capability and insufficient edge information representation. In addition, they have difficulty preserving fine details during the upsampling process. These issues cause them to perform poorly in complex pest detection tasks. To address these challenges, MEGE-DETR is proposed, a novel end-to-end pest detection model based on RT-DETR, with enhanced optimization for edge feature extraction of small insect targets. First, a new edge-enhanced backbone network, MEGFNet, is designed, incorporating a multi-scale edge generator and edge fusion module to improve the preservation of edge detail features. Subsequently, a polarity-aware linear attention encoder module is introduced to enhance the ability of the model to capture local texture and edge information. Finally, a learnable multi-band wavelet upsampling method is proposed to better retain high-frequency features during the upsampling process. Experimental results on the dataset collected from actual field scenarios involving five pest categories demonstrate that significant performance improvements are achieved by the proposed model. MEGE-DETR boosts mAP@0.5 and mAP@0.5:0.95 by 2.2% and 7.4%, respectively, while increasing detection speed to 106 frames per second, 77% faster than the baseline model. It also reduces the number of parameters and giga floating point operations to 11.9M and 39.2, which are only 37.1% and 37.9% of those in the baseline model. The results highlight the superior accuracy, efficiency and robustness of MEGE-DETR in real-world pest detection applications.
Pang et al. (Sun,) studied this question.