一种基于卷积神经网络和条件随机场的人脸检测方法

发布时间：2018-04-13 07:50

本文选题：人脸检测 + 深度学习　；参考：《华中科技大学》2016年硕士论文

【摘要】：人脸检测是一个复杂的模式判别问题,其难点主要由成像角度不同所引起:如平面内旋转和平面外旋转,偏转角度会直接影响判定人脸的准确度。当前基于深度学习卷积神经网络的检测方法虽然有着很高的检测率,但是在神经网络的输出层对人脸的处理不够精确,忽略了一张人脸对应的多个检测窗口之间的关联关系,从而导致最终人脸框不够精确。结合条件随机场模型CRF对神经网络的输出层进行调整,使得最终的人脸框更加精确。提出了一种基于卷积神经网络和条件随机场模型的人脸检测方法CRF-CNN,该方法提高了最终人脸框的精确度。方法首先对卷积神经网络进行训练,得到判定人脸和非人脸的分类器,对输入图像进行滑动窗口人脸检测,得到包含人脸的窗口;然后标注同一张人脸对应的所有检测窗口,窗口对应的置信分作为条件随机场CRF的随机变量,通过CRF模型计算窗口之间的关联关系,根据关联关系的紧密程度对窗口进行取舍;最后根据面积重叠的大小和横向距离、纵向距离重叠的大小分别对同尺度和不同尺度的窗口进行合并,得到最终的人脸框。为了使得检测率更高,该方法还对输入图片做了不同尺度的缩放处理,缩放程度的不同只会很小程度影响检测时间,不会影响检测的正确性,所以本方法对选用何种缩放算法及其参数并不敏感。实验分别与卷积神经网络检测方法DDFD、R-CNN和局部特征检测方法DPM进行了比较。结果表明,CRF-CNN的准确率和召回率与DDFD相近,高于R-CNN和DPM。在面内旋转和面外旋转的人脸检测中,CRF-CNN得到的人脸框更加准确,尤其在面外旋转的人脸检测中,CRF-CNN置信分均值为0.99759,高出DDFD 0.00527。
[Abstract]:Face detection is a complex pattern discrimination problem, which is mainly caused by different imaging angles. For example, rotation in plane and rotation out of plane, deflection angle will directly affect the accuracy of face determination.Although the current detection method based on deep learning convolution neural network has a high detection rate, the processing of face in the output layer of neural network is not accurate enough, and the correlation relationship between multiple detection windows corresponding to a face is ignored.As a result, the final face frame is not accurate enough.Combined with conditional random field model (CRF), the output layer of neural network is adjusted to make the final face frame more accurate.A face detection method CRF-CNN based on convolution neural network and conditional random field model is proposed, which improves the accuracy of the final face frame.Methods firstly, the convolutional neural network was trained to obtain the classifier for judging face and non-face, and the sliding window face detection was carried out on the input image, and the window containing the face was obtained, and then all the detection windows corresponding to the same face were labeled.The confidence score of the window is regarded as the random variable of conditional random field CRF, the correlation relation between windows is calculated by CRF model, and the window is chosen according to the tightness of the correlation relation. Finally, according to the size of area overlap and the horizontal distance,The size of vertical distance overlaps is used to merge the windows of the same scale and different scales, and the final face frame is obtained.In order to make the detection rate higher, the method also makes different scale scaling of the input image. The different scaling degree will only affect the detection time to a very small extent, and will not affect the accuracy of the detection.Therefore, this method is not sensitive to the selection of scaling algorithm and its parameters.The experiments are compared with the convolutional neural network detection method DDFDR-CNN and the local feature detection method DPM.The results showed that the accuracy and recall rate of CRF-CNN were similar to those of DDFD and higher than those of R-CNN and DPM.In the in-plane rotation and out-of-plane rotation of the face detection, CRF-CNN is more accurate, especially in the out-of-plane rotation of the face detection, the average confidence score of CRF-CNN is 0.99759, which is higher than DDFD 0.00527.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183

【参考文献】