基于生成对抗网络的图像着色技术的研究

Research on Image Coloring Techniques based on Generative Adversarial Networks

目前利用photoshop对黑白图像进行剪辑再对其进行着色是个耗费人力物力的工作,同时这种方法暴露的最大缺陷就是图像着色后的效果比较单一化。本文主要针对黑白和灰度图像上色和老照片修复工程设计优化了一种卷积神经网络CNN(Convolutional Neural Networks)的基于生成对抗网络GAN(Generative Adversarial Networks)的上色技术,可以明显提高上色速度和视觉震撼效果,并在传统方法上进行了参数和网络架构以及激活函数等内容调整。在生成器对抗器两者的互相博弈优化下,模型进行色彩学习和特征提取,获取并对模型产生的图片进行测试的同时对黑白图片上色处理。经实验测试,本文方法能在传统方法上提高拟合度和上色效果。

At present,editing and coloring black-and-white images with Photoshop is a labor-intensive and material work.At the same time,the biggest defect exposed by this method is the simplification of the effect of image coloring.In this paper,a convolutional neural networks(CNN)coloring technology is mainly designed and optimized based on generative adversarial networks(GAN)for black-and-white and gray-scale image coloring and old photo restoration engineering,which can significantly improve the coloring speed and visual shock effect,and adjusts the parameters,network architecture and activation function in the traditional methods.Under the mutual game optimization between the generator and the adversary,the model carries out color learning and feature extraction,obtains and tests the pictures generated by the model,and colors the black-and-white pictures at the same time.The experimental results show that this method can improve the fitting degree and coloring effect compared with the traditional method.