基于深度学习的多角度人脸检测方法研究

Research on Multi-angle Face Detection Method Based on Deep Learning

基于多角度的人脸检测越来越受到关注,特别是在公安领域侦破案件过程中,通过捕捉人脸图像对犯罪嫌疑人进行检测识别被广泛应用。但是在实际图像采集过程中,由于人脸姿势以及光照等环境因素的不确定性和多变性,往往会导致人脸系统无法对该类人脸进行较为精确的定位。文中基于DenseNet-201对YOLOV2算法进行了改进,提出了一种基于深度学习的多角度人脸检测方法。首先,在YOLOV2算法的基础上,使用DenseNet-201模型对人脸进行特征提取,并结合带有锚点框的卷积层在主干网络提取到的人脸特征图上进行人脸定位;然后,通过在DenseNet-201模型中的过渡层中引入归一化层使模型收敛速度加快;最后,在CelebA和FDDB人脸数据集上对YOLOV2和改进的YOLOV2方法进行测试,针对不同角度、不同光照、不同数据集对算法性能进行测试。实验结果表明,改进后的YOLOV2算法对多角度人脸检测的准确性更高,且具有更强的鲁棒性。

Face detection based on multi-angle has attracted more and more attention,especially in the process of detecting cases in the field of public security,the detection and recognition of criminal suspects by catching face images has been widely used.However,in the actual image acquisition process,due to the uncertainty and variability of environmental factors such as face posture and lighting,the face system is often unable to locate this kind of face accurately.For this,the YOLOV2 algorithm is improved based on DenseNet-201,and a multi-angle face detection method based on depth learning is proposed.First of all,on the basis of YOLOV2 algorithm,DenseNet-201 model is used for the face feature extraction,and the convolution layer with the anchor frame is combined to carry on the face localization on the face feature map extracted by the backbone network.Then,by introducing the normalization layer into the transition layer of the DenseNet-201 model,the convergence speed of the model is accelerated.Finally,the YOLOV2 and the improved YOLOV2 are tested on CelebA and FDDB face datasets.The performance of the algorithm is tested for different angles,different lighting and different datasets.The experiment shows that the improved YOLOV2 algorithm is more accurate and robust to multi-angle face detection.