Scholarlink

With the advancement of mobile device hardware capabilities and the rapid progress of artificial intelligence technologies, the focus of research has shifted towards face detection technology on mobile devices, particularly in efficiently and accurately detecting small faces characterized by low resolution and susceptibility to occlusion. This paper addresses this challenge by introducing an algorithm named CA-YOLOv8n, which integrates a lightweight intra-group attention mechanism with an adaptive feature fusion strategy to enhance the detection performance of small faces. Initially, a lightweight attention mechanism module called CRIAM is proposed to amplify features of small faces while mitigating background interference. Subsequently, an Adaptive Feature Fusion module (AFF) is developed to dynamically adjust its fusion strategy based on facial features of varying sizes, thereby increasing detection precision. Ultimately, experiments conducted on the WiderFace dataset indicate that compared to the benchmark algorithm YOLOv8n, the proposed algorithm achieves a 7.1% improvement in mAP50 and a 3.9% improvement in mAP50-95.

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

Research on Small Face Detection Algorithm Based on Lightweight Intra-Group Attention Mechanism and Multi-Scale Feature Fusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhe Yuan,
  • Jianglei Gong,
  • Baolong Guo,
  • Jiawei Song

摘要

With the advancement of mobile device hardware capabilities and the rapid progress of artificial intelligence technologies, the focus of research has shifted towards face detection technology on mobile devices, particularly in efficiently and accurately detecting small faces characterized by low resolution and susceptibility to occlusion. This paper addresses this challenge by introducing an algorithm named CA-YOLOv8n, which integrates a lightweight intra-group attention mechanism with an adaptive feature fusion strategy to enhance the detection performance of small faces. Initially, a lightweight attention mechanism module called CRIAM is proposed to amplify features of small faces while mitigating background interference. Subsequently, an Adaptive Feature Fusion module (AFF) is developed to dynamically adjust its fusion strategy based on facial features of varying sizes, thereby increasing detection precision. Ultimately, experiments conducted on the WiderFace dataset indicate that compared to the benchmark algorithm YOLOv8n, the proposed algorithm achieves a 7.1% improvement in mAP50 and a 3.9% improvement in mAP50-95.