基于单目图像的人脸深度估计

Face depth estimation based on monocular image

近年来,深度卷积神经网络在人脸识别、特征提取等图像处理任务上都展现出了极为优越的性能,基于深度网络与条件随机场的算法也在人脸深度识别的任务中取得了很好的效果。但是,随着卷积网络连续不断的下采样操作如卷积和池化,图像的分辨率也随之降低,图像细节随之丢失,十分不利于像素级的图像处理任务。因此,本文对于以往基于深度网络与条件随机场结合的算法进行了改进,利用残差连接将下采样过程中各层信息向前传递。算法将问题表述为求解条件随机场的最大化后验概率问题,并以此构建损失层,可实现整个网络端到端的训练,之后通过求解条件随机场的最大化后验概率,求得深度图像的最优解。通过实验,算法在NYU Depth V2数据集上得到验证,准确度较改进前算法具有明显提高,表明了算法的有效性。

Recently,very deep convolutional neural networks have shown outstanding performance in image processing tasks,such as face recognition and Image feature extraction.However,repeated subsampling operations like pooling or convolution striding in deep CNNs lead to decrease in the image resolution and lost of image details.In this paper,residual connection is used to pass forward the information along the down-sampling process,in order to improves the previous algorithm based on the combination of CNN and CRF.The algorithm formulates the problem as maximization of posteriori probability of conditional random fields,and constructs the loss layer based on it.In this way,the entire network can be trained end-to-end and then the MAP is obtained as the optimal inference of the depth image.Through experiments,the algorithm is tested on the NYU Depth V2 dataset,and its accuracy is obviously improved compared with the original algorithm,which shows its effectiveness.