单幅图像训练深度神经网络的编辑传播方法

Edit Propagation Using Deep Neural Network from a Single Image

针对编辑传播技术中存在的需要选择合适图像特征、手动调整图像特征权重等问题,提出一种单幅图像训练深度神经网络的编辑传播方法.首先将用户交互转换成距离图并与输入图像级联生成多通道图像,有效地结合图像的视觉和空间特征;其次以多通道图像子块作为深度神经网络的输入,抽取符合用户交互的深度特征,并对深度神经网络进行端到端的训练,从而自动分配图像特征的权重;最后将学习的网络模型作为分类器,估计图像像素属于每类用户交互的概率值,进一步后处理获得高质量的图像编辑.采用MARA 1k数据的实验结果表明,该方法能够很好地响应用户交互以进行编辑传播.

This paper proposes a novel edit propagation approach using deep neural network(DNN)from a single image,which aims to handle the problems such as appropriate features choosing and manual feature tuning.Firstly,we transform user interactions into distance maps which are then concatenated into the input image to create a new image with multiple channels,combining low-level visual features with spatial features.Secondly,we extract small multi-channel patches and use them as input of a DNN that extracts deep features adapted to user interactions.And the DNN can perform a joint end-to-end learning of visual feature and spatial feature for edit propagation,which automatically determines the importance of image features.Finally,we use the DNN as a classifier to estimate probabilities of all image pixels,and obtain editing results with high quality through further post-processing.The experimental results on MARA 1k database demonstrate that our method can respond to user interactions well and perform significantly better to propagate image edits.