An Effective Anchor-Free model with Transformer for Logo Detection Efficient Logo Detection via Transformer
Baoyan Cheng, Huichuan Duan, Sujuan Hou, Awudu Karim, Weikuan Jia, Yuanjie Zheng · 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI) · 2021
Currently, logo detection has driven research into diverse fields and services, such as video advertising, copyright infringement detection, product brand management on social media and so on. Compared to general object detection, logo detection is challenging since the real-world logo images are largely varied in their appearances and have higher complexity in their backgrounds. While many studies focus on Anchor-based solutions for object detection, few focus on Anchor-free for logo detection. In order to achieve good performance for logo detection, this paper proposes an effective Anchor-Free model with Transformer (AFT-Net), a technique that aims at recognizing logos from images. On one hand, DIoU Loss is firstly used for bounding box regression instead of IoU Loss to ensure faster convergence speed and higher accuracy. On the other hand, the Transformer Attention Mechanism is integrated into AFT-Net to enhance the representation ability of the network. The experimental results demonstrate the effectiveness of the proposed AFT-Net, it obtains 2.4% mAP improvement compared with FSAF method on Logo-Det-3K-1000. It is superior in terms of accuracy of small object recognition in logo detection compared with the state-of-the-art detection methods.