To address the challenges of small pest target sizes, dense distributions, strong background interference, and missed detection of small objects in insect pest monitoring lamp images, this study proposes YOLO11-MSC, a lightweight agricultural pest detection model based on an improved YOLO11. First, an Object-Centric Adaptive Slicing (OCAS) strategy is designed to construct the Insect35OCAS dataset, which increases the relative proportion of small pest targets in the input images while reducing redundant background interference. Second, the Multi-path Nonlinear Star Aggregation (MNS) module and the C2-based Improved Bi-Level Routing Attention (C2iBRA) module are introduced to enhance fine-grained feature extraction for multi-scale pest targets and improve background suppression in complex scenes. Finally, a Semantic-Guided Adaptive Slicing (SGAS) inference strategy is developed to transfer the sliced-image training model to full-image inference on original high-resolution insect pest monitoring lamp images. Experimental results show that YOLO11-MSC achieves an mAP@0. 5, Precision, and Recall of 95. 2%, 91. 8%, and 93. 0% on the Insect35OCAS dataset, respectively, improving the baseline YOLO11n by 1. 3, 1. 0, and 1. 2 percentage points. The model contains only 3. 4 M parameters and requires 7. 8 Giga Floating-Point Operations (GFLOPs). When combined with SGAS, YOLO11-MSC achieves an mAP@0. 5, Precision, and Recall of 86. 4%, 81. 2%, and 90. 1% on the original full-image dataset, respectively. These results demonstrate that the proposed method effectively improves the detection accuracy of small pests and the full-image inference capability in complex insect pest monitoring lamp scenarios while maintaining low model complexity, providing technical support for intelligent agricultural pest monitoring and precision control.
Li et al. (Sun,) studied this question.