MS-YOLO: Multi-Scale Network for Small-Object Pest Detection in Cotton Fields

To address the persistent challenges of detecting small-sized and scale-diverse agricultural pests in complex field environments, we propose an advanced MS-YOLO network to overcome performance limitations in crop pest monitoring. The solution features three strategically integrated components: First, the cross-stage dual-channel feature extraction block (CSDCFEB) implements coordinated spatial-channel feature extraction through cross-stage hierarchical connections, preserving fine-grained insect characteristics while minimizing information degradation during deep network propagation; Second, the C2BRA attention mechanism employs dual-branch channel-spatial recalibration to dynamically amplify discriminative regions, effectively enhancing the model's capacity to identify subtle pest patterns against cluttered backgrounds. Finally, the Multi-Scale Feature Fusion Pyramid (MSFFP) architecture establishes bidirectional multiscale fusion pathways that synergistically combine low-level textual details with high-level semantic cues, thereby reinforcing feature representation robust ness across varying pest sizes. Experimental results show that MS-YOLO achieves a mAP50 of 67.1 % on the self-built NSC dataset, an improvement of 4.3 % over the base line, while reaching 69.8 % on the large public dataset Pest 24 and 89.2 % on the small public dataset Yellow Sticky Traps. Comprehensive evaluations on these three datasets with different scales and characteristics demonstrate the significant superiority of MS-YOLO, with overall performance surpassing existing object detectors.