GBSA-YOLOv8：面向复杂田间场景稻田害虫的多尺度实时检测模型

GBSA-YOLOv8: A Multi-Scale Real-Time Detection Model for Rice Field Pests in Complex Field Scenarios

摘要: 精准农业发展背景下，基于计算机视觉的水稻害虫智能识别技术是病虫害高效防治的关键方向。受稻田光照不均、叶片遮挡、害虫体积微小等因素影响，现有模型存在小目标漏检率高、背景干扰严重、部署效率低等问题。针对YOLOv8模型局限，本文提出3阶段改进：将C2f模块升级为融合Ghost与动态卷积的C2f-Ghost-DynamicConv模块，降参提效；用BiFPN替换PANet，强化多尺度特征双向融合；嵌入SE通道注意力机制，抑制噪声并突出害虫关键特征。实验表明，改进模型精确度、召回率、平均精度分别提升3.6%、1.7%、1.9%，mAP达97.9%，有效缓解漏检误检，满足无人机田间巡检实时性需求，为水稻病虫害智能监测提供有力支撑。

Abstract: In the context of precision agriculture development, computer vision-based intelligent identification technology for rice pests has become a key direction for efficient pest control. Existing models face challenges such as high false-negative rates for small targets, severe background interference, and low deployment efficiency due to factors like uneven light distribution in rice paddies, leaf shading, and tiny pest sizes. To address the limitations of the YOLOv8 model, this paper proposes three-stage improvements: upgrading the C2f module to a C2f-Ghost-DynamicConv module that integrates Ghost and dynamic convolution for parameter reduction and efficiency enhancement; replacing PANet with BiFPN to strengthen bidirectional multi-scale feature fusion; and embedding a SE channel attention mechanism to suppress noise and highlight key pest features. Experimental results demonstrate that the improved model achieves 3.6% accuracy improvement, 1.7% recall rate increase, and 1.9% average precision gain, with mAP reaching 97.9%. This effectively reduces false negatives and false positives, meets the real-time requirements for drone field inspections, and provides robust support for intelligent rice pest monitoring.