Scholarlink

<p>Small object detection is crucial in areas such as autonomous driving and surveillance. However, it faces challenges of feature information loss and semantic insufficiency in multi-scale scenarios. To address these issues, we propose a Multi-scale Token-Channel Fusion Detection Transformer (TCF-DETR) that leverages Channel Aggregation Token Mixer (CATM) and Squeeze-Excitation Token Mixer (SETM) modules. First, multi-scale features are extracted from the backbone network and the highest semantic feature is refined through context intra-scale interactions, effectively enhancing feature representation. Second, a cross-scale fusion network based on the token-channel mixer is proposed to fuse multi-scale features. In the top-down path, CATM enhances the semantic fusion of low-level features, while SETM, used in the bottom-up path, improves the fusion of high-level features with localization information. Finally, the fused features are decoded and mapped to category labels and bounding box coordinates. Extensive experiments on the VisDrone, TinyPerson, and Cityscapes datasets shows that TCF-DETR achieves mAP50 values of 49.9%, 24.2%, and 59.5%, respectively, while maintaining a relatively low parameter count of 19.8M. The proposed model not only enhances small object detection but also shows broader applicability. The code and dataset are available at: <a href="https://github.com/Leihuan42/TCF-DETR">https://github.com/Leihuan42/TCF-DETR</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

TCF-DETR: multi-scale token-channel fusion transformer for enhanced small object detection

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Huan Lei,
  • Ze Wu,
  • Lei Shang,
  • Hong Zhao,
  • Wenyuan Yang

摘要

Small object detection is crucial in areas such as autonomous driving and surveillance. However, it faces challenges of feature information loss and semantic insufficiency in multi-scale scenarios. To address these issues, we propose a Multi-scale Token-Channel Fusion Detection Transformer (TCF-DETR) that leverages Channel Aggregation Token Mixer (CATM) and Squeeze-Excitation Token Mixer (SETM) modules. First, multi-scale features are extracted from the backbone network and the highest semantic feature is refined through context intra-scale interactions, effectively enhancing feature representation. Second, a cross-scale fusion network based on the token-channel mixer is proposed to fuse multi-scale features. In the top-down path, CATM enhances the semantic fusion of low-level features, while SETM, used in the bottom-up path, improves the fusion of high-level features with localization information. Finally, the fused features are decoded and mapped to category labels and bounding box coordinates. Extensive experiments on the VisDrone, TinyPerson, and Cityscapes datasets shows that TCF-DETR achieves mAP50 values of 49.9%, 24.2%, and 59.5%, respectively, while maintaining a relatively low parameter count of 19.8M. The proposed model not only enhances small object detection but also shows broader applicability. The code and dataset are available at: https://github.com/Leihuan42/TCF-DETR.