<p>Accurate and efficient pest detection is important for intelligent agricultural monitoring and integrated pest management, yet small pest targets, cluttered field backgrounds, dense distributions, and large pose variations remain challenging in yellow sticky-trap scenarios. To address these issues, this study proposes YOLO-DB, an improved You Only Look Once (YOLO)-based detector for small-pest recognition in agricultural imagery. In the backbone, a C2f module with deformable convolution, termed C2fDCN, is introduced to improve the adaptability of local receptive fields to irregular pest shapes and diverse postures through bounded small-magnitude offsets. In the neck, a Bidirectional Multi-Weighted Feature Pyramid Network (Bi-MW-FPN) is designed to enhance cross-scale feature fusion and strengthen high-resolution low-level features for small-object localization. YOLO-DB was evaluated on a self-collected yellow sticky-board dataset with 459 images and eight pest categories and a public forestry pest dataset with 1873 images and seven categories. On the sticky-board dataset, YOLO-DB achieved 96.5% mAP@0.5 and 64.3% mAP@0.5:0.95, outperforming YOLOv8n by 4.2 percentage points in mAP@0.5 while using only 2.29M parameters and 7.4 GFLOPs. On the public dataset, YOLO-DB achieved 98.7% mAP@0.5 and 81.5% mAP@0.5:0.95. Comparative experiments and ablation studies further demonstrate the effectiveness of the proposed C2fDCN and Bi-MW-FPN modules. These results indicate that YOLO-DB provides an efficient and lightweight solution for automated small-pest detection in agricultural monitoring scenarios.</p>