Super-resolution of face images, known as Face Hallucination (FH), has been excessively studied in recent years. Modern FH methods use deep Convolution Neural Networks (CNN) with a pixel-wise MSE loss function to infer high-resolution facial images. The MSE-oriented approaches generate over-smooth results, particularly when dealing with very low-resolution images. Recently, Generative Adversarial Networks (GANs) have successfully been exploited to synthesize perceptually more pleasant images. However, the GAN-based models do not guarantee identity preservation during face super-resolution. To address these challenges, we have proposed a novel Wavelet-integrated, Identity Preserving, Adversarial (WIPA) approach. Specifically, we present Wavelet Prediction blocks attached to a Baseline CNN network to predict wavelet missing details of facial images. The extracted wavelet coefficients are concatenated with original feature maps in different scales to recover fine details. Unlike other wavelet-based FH methods, this algorithm exploits the wavelet-enriched feature maps as complementary information to facilitate the hallucination task. We introduce a wavelet prediction loss to push the network to generate wavelet coefficients. In addition to the wavelet-domain cost function, a combination of perceptual, adversarial, and identity loss functions has been utilized to achieve low-distortion and perceptually high-quality images while maintaining identity. The extensive experiments prove the superiority of the proposed approach over the state-of-the-art methods by achieving PSNR of 25.16 dB for CelebA dataset and verification rate of 86.1% for LFW dataset; both conducted on 8X magnification factor.

近年来，人们对人脸图像的超分辨率，即人脸幻觉（FH）进行了过度研究。现代 FH 方法使用具有逐像素 MSE 损失函数的深度卷积神经网络 (CNN) 来推断高分辨率面部图像。面向 MSE 的方法会产生过度平滑的结果，尤其是在处理分辨率非常低的图像时。最近，生成对抗网络 (GAN) 已成功用于合成感知上更令人愉悦的图像。然而，基于 GAN 的模型不能保证在人脸超分辨率期间保持身份。为了应对这些挑战，我们提出了一种新颖的小波集成、身份保持、对抗（WIPA）方法。具体来说，我们提出了附加到基线 CNN 网络的小波预测块，以预测面部图像的小波缺失细节。提取的小波系数与不同尺度的原始特征图连接以恢复精细细节。与其他基于小波的 FH 方法不同，该算法利用小波丰富的特征图作为补充信息来促进幻觉任务。我们引入了小波预测损失来推动网络生成小波系数。除了小波域成本函数之外，还利用感知、对抗和身份损失函数的组合来实现低失真和感知高质量的图像，同时保持身份。大量实验证明了所提出的方法优于最先进的方法，CelebA 数据集的 PSNR 为 25.16 dB，LFW 数据集的验证率为 86.1%；均在 8X 放大倍数下进行。