基于多任务深度学习的快速人像自动抠图

Fast portrait automatic matting based on multi-task deep learning

针对大多数人像抠图存在时效性低、需要人工标注三分图和依赖颜色作为主要依据而导致精度难提高的问题,提出一种基于多任务学习的神经网络的快速人像自动抠图算法。该方法首先对图像的三分图进行学习预测,并将得到的信息反馈至网络后去学习预测图像的α值,网络结构采用编码器-解码器的方式,编码器部分使用深度可分离残差卷积做特征提取,和多组空洞卷积并联组合使得网络拥有足够的感受野;解码部分使用双线性加性上采样使特征图逐步恢复至原图大小,另外使用跳跃连接层将编码和解码部分相连接;使用公开数据库作为测试集,与Deeplab加文献[3]算法和DAPM算法进行对比,实验结果表明在运行时间、SAD、MSE和Gradient评价指标上优于对比算法。

As the existing portrait matting techniques have the shortcomings of time-consuming,manual marking of trimap,and difficult to improve the accuracy due to the dependence on color as the main basis,a novel fast portrait automatic matting method based on multi-task deep learning is proposed.Firstly,the trimap of the image is learned to predict,and the obtained information is fed back to the network to learn to predict the value ofαof the image.The network structure of encoder-decoder is adopted in this method.The depth separable residual convolution for feature extraction and the combination of multiple sets of dilated convolutions are used in encoding part to make the network have enough receptive field.The decoding part uses bilinear additive for up-sampling to restore the feature graph to its original size.Skip connections are used to connect the encoding and decoding parts.Compared with algorithms of Deeplab+IMFM and DAPM,experimental results show that the evaluation index of the method proposed is superior to the contrast algorithm in terms of running time,SAD,MSE and Gradient,by using public databases as test set.