Scholarlink

<p>Object detection is a fundamental task in computer vision. It allows machines to locate and recognize objects in digital images. Convolutional neural networks (CNNs) reduce the spatial size of feature maps as the network deepens. Large objects are mainly represented in deeper layers, while details of small objects are kept in shallower layers. Standard feature fusion methods like the Feature Pyramid Network (FPN) struggle to combine these features effectively. This work aims to improve multi-scale feature fusion for better object detection performance. We propose a new Complementary Feature Pyramid Network (CFPN). It starts by fusing features from adjacent layers and then gradually includes non-adjacent layers to enrich the representation. A residual connection is added at each level to keep the original information and prevent it from being overwritten. We integrated CFPN into the Faster R-CNN framework and evaluated it on the MS COCO 2017 dataset. Our method achieved an average precision (AP) of 39.8% on the validation set, which is 2.3% higher than the result of the original FPN. These results show that better multi-scale feature fusion can improve object detection.</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

Enhanced object detection via complementary feature pyramid networks

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Li Zhang,
  • Chen Xu,
  • KaiBing Zhang,
  • GuangDong Zhu,
  • Ting Yan,
  • MeiLing Lu

摘要

Object detection is a fundamental task in computer vision. It allows machines to locate and recognize objects in digital images. Convolutional neural networks (CNNs) reduce the spatial size of feature maps as the network deepens. Large objects are mainly represented in deeper layers, while details of small objects are kept in shallower layers. Standard feature fusion methods like the Feature Pyramid Network (FPN) struggle to combine these features effectively. This work aims to improve multi-scale feature fusion for better object detection performance. We propose a new Complementary Feature Pyramid Network (CFPN). It starts by fusing features from adjacent layers and then gradually includes non-adjacent layers to enrich the representation. A residual connection is added at each level to keep the original information and prevent it from being overwritten. We integrated CFPN into the Faster R-CNN framework and evaluated it on the MS COCO 2017 dataset. Our method achieved an average precision (AP) of 39.8% on the validation set, which is 2.3% higher than the result of the original FPN. These results show that better multi-scale feature fusion can improve object detection.