Towards Human-like Performance Face Detection : A Convolutional Neural Network Approach
Towards Human-like Performance Face Detection : A Convolutional Neural Network Approach
复制标题
迈向类人性能人脸检测：卷积神经网络方法
DOI：
--
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
J. Kleef
中科院分区：
文献类型：
--
作者：
J. Kleef
人脸检测是计算机视觉领域中一个被广泛讨论的课题。人脸检测的应用范围从方便的HMI到政府规模的监控应用。当提供的图像处于完美状态（例如官方肖像）时，面部检测器表现良好。在这种情况下，面部与相机的透镜完全对准，没有遮挡，并且没有姿势。人脸检测器面临的挑战是当人脸不处于完美状态时进行检测，这种情况在真实的生活中经常发生。这项研究的目标是找出为基于卷积神经网络（CNN）的人脸检测器提供更多训练数据的影响。人脸检测器由深度学习框架Caffe构建。用于人脸检测的模型是AlexNet。创建了10个版本的人脸检测器，大小不同，经过了多少次迭代。在具有最少和最多训练数据的版本之间，从39到64个正确接受的性能增益，因此当呈现更多训练数据时，性能会增加。
Face detection is a widely discussed topic in the field of computer vision. The application of face detection varies from convenient HMIto government-scale surveillance applications. The face detectors perform well when a image is provided that is in a perfect condition, such as an official portrait. In this condition a face is perfectly aligned to the lens of the camera, not occluded and it has no pose. The challenge for a face detector is to detect a face when it is not in a perfect condition and this situation happens more often than not in real life-situations. The goal of this research is to find out what the impact of providing more training data to a convolutional neural network(CNN)based face detector is. The face detector is constructed by the deep learning framework Caffe. The model used for the face detection is AlexNet. There were 10 versions of face detectors created, varying in size and how many iterations it gone through. There is a performance gain from 39 to 64 correct accepts, between the versions with the least and most training data, so the performance is increased when more training data is presented.