A DenseUnet Generative Adversarial Network (GAN) is proposed to colorize near-infrared (NIR) face images. The GAN generator incorporates the advantages of both DenseNet and Unet structures. The DenseNet extracts facial features effectively by increasing the network depth, and the Unet keeps important facial details through skip-connection. The generator further integrates an optimized facial loss function designed by considering pixel loss, color loss and feature loss. The proposed network is evaluated against state-of-the-art grayscale image colorization GANs over two separate datasets. Both qualitative and quantitative comparisons demonstrate that with improved network structures and loss constrains the DenseUnet GAN can colorize NIR face images with natural color, minimal face shape distortion, and rich facial details.

提出了DenseUnet生成对抗网络（GAN）对近红外（NIR）人脸图像进行着色。GAN生成器结合了DenseNet和Unet结构的优点。DenseNet通过增加网络深度来有效提取面部特征，而Unet通过跳过连接保留重要的面部细节。该生成器还集成了优化的面部损失功能，该功能通过考虑像素损失，颜色损失和特征损失而设计。在两个独立的数据集上，针对最新的灰度图像着色GAN评估了所建议的网络。定性和定量比较均表明，通过改进的网络结构和损耗约束，DenseUnet GAN可以使NIR面部图像着色，具有自然色彩，最小的面部形状失真和丰富的面部细节。