基于双重约束CNN的壁画图像色彩还原

Color restoration of mural image based on double constrained convolutional neural network

针对石窟壁画图像普遍存在的褪变色问题,提出了一种基于双重约束卷积神经网络的壁画图像色彩还原方法.该方法采用卷积神经网络提取壁画图像的高阶特征,在此基础上,为了对壁画色彩进行全局还原,利用最大均值差异约束建立全局特征映射关系以还原壁画图像中的全局颜色特征.针对不同卷积层存在局部色彩信息缺失的问题,利用马尔科夫随机场约束进行空间布局限制,保留壁画图像中的局部结构信息并对局部色彩进行还原.此外,通过平方梯度约束抑制图像还原过程中的噪声.实验结果表明:该方法在主客观分析指标方面均表现较佳,对于壁画褪变色区域的整体还原效果较好.

Aiming at the problem of discoloration in grotto mural images,a color restoration method for mural images based on double constrained convolutional neural network was proposed.A convolutional neural network was used to extract the higher-order features of the mural image.Based on this,in order to carry out global restoration of mural color,the maximum mean difference constraint was used to establish a global feature mapping relationship to restore the global color feature in mural image.For the lack of local color information in different convolutional layers,Markov random field constraint was used to restrict the spatial layout.This constraint can retain the local structure information of the mural image and restore the local color of mural image.In addition,the square gradient constraint was used to suppress the noise in the image restoration process.The experimental results show that this method is better in both subjective and objective indicators and has a better overall restoration effect in the faded or discolored areas of mural images.