Abstract: In order to meet the fast and accurate detection requirements for fuzzy flower images in complex environments,a composite model SR⁃YOLOv8n⁃BCG was proposed.This model effectively integrated the image reconstruction capability of SRGAN(super⁃resolution generative adversarial network)and the object detection capability of YOLOv8.Furthermore，the network structure was further improved to enhance accuracy and achieve lightweight design.Firstly,SR⁃YOLOv8n⁃BCG employed SRGAN to perform super⁃resolution processing on fuzzy flower images,enhancing the image quality input into model.Secondly,in the feature extraction network of YOLOv8n,a weighted bidirectional feature pyramid network(BiFPN)was used to replace the PAN⁃FPN module,effectively fusing multi⁃scale flower features and reducing the model size.Additionally,a coordinate attention(CA)mechanism was introduced to enhance the model’s feature extraction ability.Finally,Ghost convolution was used to replace regular convolution,further improving detection accuracy and model lightweightness.Experimental results on a self⁃made dataset of five flower categories demonstrated that compared to SR⁃YOLOv8n,the SR⁃YOLOv8n⁃BCG model achieved an average precision improvement of 1.2 percentage points,reaching 95.4% with a 35.5% reduction in model size.The results above indicate that the proposed improved model effectively enhances the detection accuracy of fuzzy flower images and achieves lightweight design to suit lower⁃end devices.

Key words: Flower, Image detection, SRGAN, YOLOv8, BiFPN, Coordinate attention mechanism, Ghost convolution

摘要： 为满足复杂环境下对模糊花卉图像快速、精确的检测需求，提出了一种组合模型SR-YOLOv8n-BCG，该模型有效地综合了SRGAN（Super⁃resolution generative adversarial network）的图像重建能力和YOLOv8 的目标检测能力，并针对网络结构进一步改进，以提高准确率并实现轻量化。首先，SRYOLOv8n-BCG 利用SRGAN 对模糊花卉图像进行超分辨处理，提高输入模型的图像质量。其次，在YOLOv8n特征提取网络中使用加权双向特征金字塔网络（Bidirectional feature pyramid network，BiFPN）替换PAN-FPN模块，以有效融合多尺度的花卉特征，并降低模型的体积。同时，引入坐标注意力机制（Coordinate attention，CA）以增强模型的特征提取能力。最后，利用Ghost卷积替换普通卷积，进一步提升检测精度并轻量化模型。结果表明，经过在自制的5 类花卉数据集上的试验评估，相较于SRYOLOv8n，SR-YOLOv8n-BCG模型在模型尺寸减小35.5%的情况下，平均精度均值提高1.2百分点，达到95.4%。表明提出的改进模型有效地提高了对模糊花卉图像检测的准确率，并实现了轻量化以适应低配的设备。

关键词: 花卉, 图像检测, SRGAN, YOLOv8, BiFPN, 坐标注意力机制, Ghost卷积