基于像素级生成对抗网络的复杂场景灰度图像彩色化

针对当前基于深度学习的彩色化模型在面对具有多个目标的复杂场景时存在的误着色问题,提出一种基于像素级生成对抗网络的彩色化模型.该模型在生成网络中采用全卷积网络模型处理不定尺度的输入灰度图像,并加入与真实彩色分量间的L1损失作为彩色化优化目标;在判别网络中,采用语义分割网络计算像素级Softmax损失,反向传递优化彩色化生成网络.在Pascal Segmentation及ILSVRC2012数据集上进行的彩色化图像质量比较,实验结果表明,与同类模型相比,本文模型在处理复杂场景灰度图像的彩色化任务中具有更高的着色准确率,并且对不同目标之间具有更好的区分度.

Colorization of Complex Scene Grayscale Images with Pixel-Wise Generative Adversarial Networks

Traditional deep learning based colorization models may cause mistaken coloring in dealing with complex scenarios.For this problem,we proposed a pixel-wise generative adversarial network based colorization method.Firstly,we built a fully convolutional network for the generative model to deal with grayscale images of uncertainty scale.Moreover,the L1 loss between the output color maps and the real color components was calculated as the optimization goal.Secondly,we utilized a semantic segmentation network to build the discriminative model,of which a pixel-wise Softmax loss was calculated and propagated back to improve the performance of the colorization model for a better coloring output.Experimental results of color image quality comparison on Pascal Segmentation and ILSVRC2012 datasets show that the proposed colorization model achieves a higher accuracy and better discrimination between different objects compared with other colorization models.