Scholarlink

<p>Despite significant advancements in face detection in general environments, performance remains inadequate in challenging scenarios, particularly when dealing with complex face detection scenes characterized by small sizes, low resolution, and varying poses. To address this issue, we propose the Face Density Augmentation Module (FDAM), which enhances feature representation and effectively mitigates background interference through a density estimation task. FDAM consists of two core components: Hierarchical Density Estimation (HDE) and Density Feature Fusion (DFF). HDE introduces multi-scale density constraints to capture fine-grained information across different levels, while DFF seamlessly integrates these density features into the detection process. Extensive experiments demonstrate that our method achieves substantial performance improvements on both the WIDER FACE and FDDB datasets. Notably, FDAM achieves these enhancements without requiring additional data or manual annotations and can be readily integrated into existing detection frameworks. The primary contributions of this paper are as follows: Firstly, we propose an innovative approach to enhance face detection based on density estimation. Secondly, we design the HDE and DFF modules to achieve effective multi-scale density estimation and feature fusion, and we validate the efficacy of integrating the density estimation task into multi-task learning frameworks. Finally, we confirm the superior performance of our method across multiple challenging datasets.</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

Multi-scale count-task guided feature enhancement face detection

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Ziyang Wu,
  • Yin Lin,
  • Qidong Huang,
  • Wengang Zhou,
  • Houqiang Li

摘要

Despite significant advancements in face detection in general environments, performance remains inadequate in challenging scenarios, particularly when dealing with complex face detection scenes characterized by small sizes, low resolution, and varying poses. To address this issue, we propose the Face Density Augmentation Module (FDAM), which enhances feature representation and effectively mitigates background interference through a density estimation task. FDAM consists of two core components: Hierarchical Density Estimation (HDE) and Density Feature Fusion (DFF). HDE introduces multi-scale density constraints to capture fine-grained information across different levels, while DFF seamlessly integrates these density features into the detection process. Extensive experiments demonstrate that our method achieves substantial performance improvements on both the WIDER FACE and FDDB datasets. Notably, FDAM achieves these enhancements without requiring additional data or manual annotations and can be readily integrated into existing detection frameworks. The primary contributions of this paper are as follows: Firstly, we propose an innovative approach to enhance face detection based on density estimation. Secondly, we design the HDE and DFF modules to achieve effective multi-scale density estimation and feature fusion, and we validate the efficacy of integrating the density estimation task into multi-task learning frameworks. Finally, we confirm the superior performance of our method across multiple challenging datasets.