基于YOLOv5改进的粘蝇纸家蝇识别算法

Enhanced Sticky Fly Paper Housefly Recognition Algorithm Based on YOLOv5

摘要: 文章针对粘蝇纸上家蝇目标体积小、形态多变且背景复杂的检测难题，提出了一种基于YOLOv5改进的检测算法。为充分反映实际应用场景，本研究搜集和采集并标注了500张粘蝇纸图像，其中400张用于训练，100张用于测试。改进工作主要集中在检测头部分，通过引入Zoom_cat模块实现多尺度特征的对齐与融合、采用ScalSeq模块增强特征序列化处理能力，并结合注意力机制提升目标区域的显著性，从而优化小目标的特征提取和定位效果。实验结果表明，改进后的模型在mAP、精确率和召回率等关键指标上均显著优于原始YOLOv5m模型，充分验证了所提方法在家蝇检测中的有效性和鲁棒性。该研究为粘蝇纸上家蝇的数量监测提供了一种高效、准确的识别方法，同时为小目标检测问题的进一步探索提供了新的思路。

Abstract: In this paper, we propose an enhanced detection algorithm, termed YOLOv5, for identifying housefly targets on sticky fly paper. These targets exhibit characteristics such as small size, variable morphology, and complex background. To thoroughly reflect the actual application scenario, a comprehensive data set was collected, acquired, and labeled. This included 500 images of sticky fly paper, with 400 utilized for training and 100 for testing purposes. The primary focus of the improvement work is on the detection head part, which optimizes the feature extraction and localization effect of small targets by introducing the Zoom_cat module to achieve multi-scale feature alignment and fusion, adopting the ScalSeq module to enhance the feature serialization processing capability, and combining with the attention mechanism to enhance the saliency of the target region. The experimental results demonstrate that the enhanced model significantly outperforms the original YOLOv5m model in terms of key metrics such as mAP, precision, and recall, thereby validating the efficacy and reliability of the proposed method for housefly detection. This study proposes an efficient and accurate identification method for the population monitoring of houseflies on sticky fly paper and concomitantly offers a novel approach for further exploration of the small target detection problem.