计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (20): 233-243.DOI: 10.3778/j.issn.1002-8331.2305-0513

• 图形图像处理 • 上一篇    下一篇

结合YOLO-FGE网络的商标检测与分类

缪春沅,王修晖   

  1. 中国计量大学 信息工程学院 浙江省电磁波信息技术与计量检测重点实验室,杭州 310018
  • 出版日期:2024-10-15 发布日期:2024-10-15

Trademark Detection and Classification Based on YOLO-FGE

MIAO Chunyuan, WANG Xiuhui   

  1. Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, College of Information Engineering, China Jiliang University, Hangzhou 310018, China
  • Online:2024-10-15 Published:2024-10-15

摘要: 为了解决商标样式众多、背景复杂、尺度变化大等问题,基于YOLOv5框架,提出了一种YOLO-FGE网络模型,以更精确地分辨出商标类别信息。提出一种新的特征增强模块来提升特征层对不同类型商标的适应性,使网络更多关注待检测商标的有用信息。在YOLOv5的C3模块中嵌入全局注意力模块对骨干网络和颈网络进行优化。提出了一种增强空间注意力模块,利用空洞卷积扩大感受野,并结合通道注意力和Transformer模块来提升商标检测精度。在图形类商标数据集上的实验结果表明,该模型将mAP提升至92.3%,比大多数现有方法具有更高的检测精度。

关键词: 商标检测, 特征增强, 全局注意力, 空间注意力

Abstract: In order to solve the trademarks’ problems about their numerous styles, complex backgrounds, and large-scale changes, a YOLO-FGE network model based on the YOLOv5 framework is proposed to distinguish trademark category information more accurately. Firstly, a feature enhancement module is put forward to enhance the adaptability of the feature layer to different kinds of trademarks, making the network pay more attention to the useful information of trademarks to be detected. Secondly, the global information attention module is embedded in the C3 module of YOLOv5 to optimize the backbone and neck network. Finally, an enhanced spatial attention module is raised, which uses dilated convolution to expand the receptive field, combines channel attention and Transformer module to improve the detection accuracy. The experimental results on the graphic trademark dataset show that the model improves mAP to 92.3%, which has higher detection accuracy than most existing methods.

Key words: trademark detection, feature enhancement, global attention, spatial attention