一种改进的生成对抗网络的图像上色方法研究-《计算机技术与发展》

文章信息/Info

Title:: Research on an Improved Method of Generative Adversarial Networks Image Coloring

Author(s):: JIANG Wen-jie; LUO Xiao-shu^* ; DAI Qin-xuan; School of Electronic Engineering,Guangxi Normal University,Guilin 541004,China

Keywords:: black and white photo / image; coloring; convolutional neural network; unsupervised learning; generative adversarial network(GAN)

摘要:: 利用手持云台对黑白照/图像进行拍摄再对其进行上色是耗时耗力的工作, 为了提高对黑白照片/图像上色的效率和视觉效果, 利用深度学习的卷积神经网络提出了一种基于改进生成对抗网络的上色算法。采用了原模型 pix2pix(image-to-image translation with conditional generative adversarial networks) 的 U 型结构的生成器并在其中引入了自注意力机制提高输出图像的色彩多样性;其次是使用实例归一化对网络结构作进一步的优化处理,生成器和判别器在网络训练期间不断进行相互对抗学习,同时模型不断学习并优化黑白照片/图像到对应彩色图像的映射关系; 最后实现了对黑白照片/图像的自动化上色,同时使用客观量化指标(MSE、PSNR 和 SSIM) 对实验结果进行评价。实验结果表明:该算法能快速有效实现对黑白照片/ 图像的无监督上色,同时保持良好的视觉效果。

Abstract:: It is time-consuming and labor-intensive to take a black and white photo/image and then paint it with a hand holder. In order to improve the efficiency and visual effect of black-and-white photo/image, we propose a coloring algorithm based on improved generative adversarial networks with deep learning convolutional neural network. The U-shaped structure of the original model pix2pix (image-to-image translation with conditional adversarial networks) is adopted and a self-attention mechanism is introduced to improve the color diversity of the output image. Followed by the use of instance normalization to further optimize the network structure, the generator and discriminator continue to compete against each other during network training, while the model continuously learns and optimizes the mapping of black and white photo/image to corresponding color images. Finally the automatic coloring of black and white photo/image is realized, and the experimental results using objective quantitative indicators (MSE, PSNR and SSIM) are evaluated, which show that the proposed algorithm can quickly and effectively achieve the unsupervised coloring of black and white photo/image while maintaining better visuals.