Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (19): 250-256.DOI: 10.3778/j.issn.1002-8331.2103-0052

• Graphics and Image Processing • Previous Articles     Next Articles

Robust Face Detection Using YOLOv3 Fusion Super Resolution Reconstruction

ZHAO Junyan, JIANG Ailian, QIANG Yan   

  1. School of Information and Computer, Taiyuan University of Technology, Jinzhong, Shanxi 030600, China
  • Online:2022-10-01 Published:2022-10-01

YOLOv3融合图像超分辨率重建的鲁棒人脸检测

赵军艳,降爱莲,强彦   

  1. 太原理工大学 信息与计算机学院,山西 晋中 030600

Abstract: Due to the face detection in complex scenes is affected by image quality, face scale, light and other factors, it is a very challenging task to accurately locate small faces and avoid missing and false detection. This paper proposes a two-level face detection model, SR-Yolov3, which based on YOLOv3 and fusion of image super resolution reconstruction technology. In view of the missing detection problem of small-scale faces in scenes, the K-means++ algorithm is used to carry out clustering analysis on the anchor boxes, and smaller anchor boxes are set to capture the information of small faces. Aiming at the problem of false detection of fuzzy small-scale faces, the Darknet53 is used as the backbone network, and the SRGAN image super-resolution reconstruction module is integrated to enhance the data of low-resolution faces, forming a detection network that can improve the detection performance of low-resolution small faces. The WIDERFACE dataset is used to train and test the SR-YOLOv3 model. And compared with MTCNN, CMS-RCNN, HR and S3FD algorithms, it is verified that the proposed model has higher detection precision, especially the performance improvement on the hard set is the most obvious. SR-YOLOv3 can effectively use face information to accurately detect hard-to-detect face targets in complex scenes, with good robustness.

Key words: face detection, YOLOv3, super resolution, convolutional neural network

摘要: 复杂场景中的人脸检测由于受到图像质量、人脸尺度和光线等因素影响,精准地定位小人脸、避免漏检、误检是一件极具挑战性的任务。提出了一种基于YOLOv3、融合图像超分辨率重建技术的两级人脸检测模型SR-YOLOv3。针对场景中小人脸目标的漏检问题,利用K-means++算法对先验框进行聚类分析,设置更小尺寸的先验框来捕获小人脸信息;针对模糊小尺度人脸的误检问题,采用Darknet53作为主干网络,融入SRGAN图像超分辨率重建模块对低分辨率的人脸进行数据增强,形成一个可以提高低分辨率小人脸检测性能的检测网络。利用WIDERFACE数据集对SR-YOLOv3模型进行训练和测试,并与MTCNN、CMS-RCNN、HR、S3FD算法相比,验证了提出的模型具有更高的检测精确度,尤其是在hard子集上的性能提升最为明显。SR-YOLOv3能够有效地利用人脸信息,精准检测出复杂场景中的难检测人脸目标,具有很好的鲁棒性。

关键词: 人脸检测, YOLOv3, 图像超分辨率, 卷积神经网络