基于深度学习的灰度图像实际颜色预测

Actual color prediction of gray images based on deep learning

针对传统色彩还原方法依赖专家辅助,在还原过程中花费大量的时间且费用高昂的问题,基于深度学习技术对该问题进行了研究与改进,提出了一种全自动的两阶段式灰度图像着色算法。首先结合分类网络和采样上色网络,并使它们共享部分相同的网络结构和权值,然后将平均平方误差和交叉熵函数的加权作为损失函数,最后在大规模场景分类数据库ImageNet上对类标和色彩进行重平衡后进行训练。实验表明,该算法输出的彩色图像更加真实、准确且色彩鲜艳,同时速度上优于传统方法。该技术可用于保证图像语义正确的情况下,将灰度图像转换为较真实的彩色图像。

Traditional color restoration method relies on expert assistance and spends a lot of time and cost in the restoration process.This problem was researched and improved based on deep learning technology,and a fully automatic two-stage grayscale image coloring algorithm was proposed.Firstly,a classification network and a sampling color network were combined,sharing some of the same network structure and weights,then the weighted value of average squared error and cross entropy function was used as the loss function,and finally the large-scale scene classification database ImageNet was used to train the algorithm after balancing with color.Experiments show that the color image output by the algorithm proposed in this paper is more realistic,accurate and colorful,and the speed is better than the traditional method.This technique can be used to convert grayscale images into more realistic color images when the image semantics are correct.