Scholarlink

<p>As e-commerce continues to expand at a steady pace, there has been an increasing rise in instances of trademark infringement. This escalation underscores the necessity for the automated detection and classification of trademarks. This paper proposes an automatic trademark detection network based on the YOLO (You Only Look Once) framework, denoted as YOLO-EES, which is designed to detect and classify trademarks within input images. Firstly, based on a dilated convolution of multiple expansion rates to obtain different receptive fields, the enhanced spatial pyramid pooling module is designed. In this module, a hybrid attention mechanism is built to fuse multiscale feature information and improve the ability of the model to detect trademarks. Secondly, an efficient self-attention module is constructed to reduce information loss during the feature map generation process and improve the feature extraction capability of the network model. Besides, the SIoU Loss is employed as the bounding box regression loss function to improve the speed and accuracy of the bounding box regression. Our experimental results based on the public dataset FlickrSportLogos-10-master and the self-constructed Trademark Content Classification Dataset show that … our model outperforms traditional methods (e.g., Faster R-CNN, YOLOv5) by 9.3–12.7% in mAP. The codes and datasets used in this study is publicly available at <a href="https://gitee.com/gait-world_0/tm-yolo-ees">https://gitee.com/gait-world_0/tm-yolo-ees</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Trademark detection and classification from digital images with self-attention mechanism

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Chunyuan Miao,
  • Xiuhui Wang,
  • Wei Qi Yan

摘要

As e-commerce continues to expand at a steady pace, there has been an increasing rise in instances of trademark infringement. This escalation underscores the necessity for the automated detection and classification of trademarks. This paper proposes an automatic trademark detection network based on the YOLO (You Only Look Once) framework, denoted as YOLO-EES, which is designed to detect and classify trademarks within input images. Firstly, based on a dilated convolution of multiple expansion rates to obtain different receptive fields, the enhanced spatial pyramid pooling module is designed. In this module, a hybrid attention mechanism is built to fuse multiscale feature information and improve the ability of the model to detect trademarks. Secondly, an efficient self-attention module is constructed to reduce information loss during the feature map generation process and improve the feature extraction capability of the network model. Besides, the SIoU Loss is employed as the bounding box regression loss function to improve the speed and accuracy of the bounding box regression. Our experimental results based on the public dataset FlickrSportLogos-10-master and the self-constructed Trademark Content Classification Dataset show that … our model outperforms traditional methods (e.g., Faster R-CNN, YOLOv5) by 9.3–12.7% in mAP. The codes and datasets used in this study is publicly available at https://gitee.com/gait-world_0/tm-yolo-ees.