基于级联卷积神经网络的人脸检测算法

Face detection algorithm based on cascaded convolutional neural network

为了解决大部分基于深度学习的方法直接提取深度抽象特征,无法在速度与精度上取得均衡问题,该文将传统的级联框架与深度卷积神经网络结合,提出了一种新的基于级联的由浅至深的卷积神经网络人脸检测方法。首先通过融合全脸与部分人脸的全卷积神经网络置信图谱快速定位人脸候选区域,然后采用深度神经网络提取人脸鲁棒性特征,对候选区域进一步分类验证,并用联合回归的方法确定最终人脸位置,提高检测精确度。所提出的方法与一些代表性的算法对比和分析,在FDDB、AFW权威评测集上达到了可比较的精度,且能快速地进行检测。

To overcome the problem that most of the methods based on depth learning cannot get a balance between speed and accuracy when directly extracting the depth of abstract features. This paper proposes a face detection method based on cascade neural network by combining traditional cascade framework with from-shallow-to-deep convolutional neural network. Firstly,this paper selects candidate face regions by means of fusing the confidence maps of images with part and full faces based on full convolutional neural network. Secondly,this paper extracts robust features of face to validate the candidate regions. Simultaneously,this paper locates the face with combined regression to improve the detection accuracy. In the experiments,the proposed method achieves comparable or better accuracy and speed on FDDB,AFW benchmarks.