| 摘要: |
| 自然场景中的实体标志, 如商标、交通标志等, 易受拍摄角度、所依附物体形变、尺度变化等影响, 导致检测精度降低. 为此, 提出一种注意力引导的标志检测与识别网络(attention guided logo detection and recognition network, AGLDN), 联合优化模型对多尺度变化和复杂形变的鲁棒性. 首先通过标志模板图像搜集及掩码生成、标志背景图像选取和标志图像生成创建标志合成数据集; 然后基于RetinaNet和FPN提取多尺度特征并形成高级语义特征映射; 最后利用注意力机制引导网络关注标志区域, 克服目标变形对特征鲁棒性的影响, 实现标志检测与识别. 实验结果表明, 所提方法可以有效降低尺度变化、非刚性形变的影响, 提高标志检测准确率. |
| 关键词: 标志检测和识别 数据合成 多尺度特征融合 注意力引导 |
| DOI：10.13328/j.cnki.jos.007033 |
| 分类号:TP391 |
| 基金项目:国家重点研发计划(2021YFF0901600); 国家自然科学基金(61672495, 61971016); 北京市自然科学基金-市教委联合资助项目(KZ201910005007) |
|
| Attention Guided Logo Detection and Recognition |
|
ZHANG Dong-Ming1, JIN Guo-Qing1, LU Ding-Yu2, ZHANG Jing2, ZHANG Yong-Dong1,3
|
|
1.State Key Laboratory of Communication Content Cognition (People’s Daily Online Co. Ltd.), Beijing 100733, China;2.Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China;3.School of Information Science Technology, University of Science and Technology of China, Hefei 230026, China
|
| Abstract: |
| In natural scenes, logos such as trademarks and traffic signs are susceptible to shooting angle, carrier deformation, and scale changes, which reduces logo detection accuracy. Thus, this study proposes an attention guided logo detection and recognition network (AGLDN) to jointly optimize the model robustness for multi-scale and complex deformation. First, a logo synthesis dataset is established by image collection and mask generation of logo templates, image selection of logo background, and logo image generation. Then, based on RetinaNet and FPN, multi-scale features are extracted and high-level semantic feature mapping is formed. Finally, the attention mechanism guided network is employed to focus on the logo area, and the influence of logo deformation on feature robustness is suppressed to improve logo detection and recognition. Experimental results show that the proposed method can reduce the influence of scale changes and non-rigid deformation, and improve detection accuracy. |
| Key words: logo detection and recognition data synthesis multi-scale features fusion attention guidance |
