计算机技术与发展

由于自然场景下粮田害虫检测任务易受目标重叠、尺度不一和复杂背景问题的影响，使得模型对关键特征的感知能力不足并导致其泛化能力下降。为此，该文对YOLOv8模型进行了改进，提出了一种基于YOLO-RDHL的粮田害虫检测方法。首先，在颈部网络引入改进后的HC-DySample模块，增强了模型对害虫目标特征的捕获精度和重叠区域感知能力。其次，在主干网络和颈部网络的C2F模块中加入并行的多分支卷积模块(DBB),提高了模型对多尺度特征的表达能力。再次，在主干网络中引入改进后的RAMLCA自注意力机制，增强了模型在复杂背景下对目标特征的识别能力。最后，将检测头网络替换为改进后的LSCDv2模块，有效提高了模型的检测速度。实验结果表明，YOLO-RDHL模型相比YOLOv8模型在Precision、Recall、mAP50、mAP50-95指标上分别提升了5.1百分点、4.4百分点、4.8百分点、1.7百分点。在推理时间和占用内存方面，YOLO-RDHL与YOLOv8模型相比无明显差距。YOLO-RDHL在保持较低推理时间和占用内存的同时，显著提升了检测精度，体现出更优的性能平衡。本研究为自然场景下粮田害虫识别和预警提供了理论和实践支撑。

Due to challenges such as target overlap, scale variation, and complex backgrounds in natural field environments, pest detection in grain fields often suffers from insufficient perception of key features, leading to a decline in model generalization ability. To address this issue, we propose an improved YOLOv8-based detection method named YOLO-RDHL for grain field pest detection. Firstly, an enhanced HC-DySample module is introduced into the neck network to improve the model's ability to capture pest features and perceive overlapping regions. Secondly, parallel multi-branch convolution modules(DBB) are integrated into the C2F modules of both the backbone and neck networks, enhancing the model's capability to represent multi-scale features. Thirdly, an improved RAMLCA self-attention mechanism is embedded in the backbone network to strengthen the model's recognition of target features in complex backgrounds. Finally, the detection head is replaced with an improved LSCDv2 module, effectively increasing the detection speed. Experimental results show that compared to the original YOLOv8 model, YOLO-RDHL achieves improvements of 5.1 percentage points, 4.4 percentage points, 4.8 percentage points, and 1.7 percentage points in Precision, Recall, mAP50,and mAP50-95,respectively. In terms of inference time and memory usage, YOLO-RDHL exhibits no significant difference compared to YOLOv8. YOLO-RDHL significantly improves detection accuracy while maintaining low inference latency and memory usage, demonstrating a better performance balance. This study provides both theoretical and practical support for pest recognition and early warning in natural field environments.