基于图像块相似性和补全生成的人脸复原算法

图像获取过程中,由于成像距离、成像设备分辨率等因素的限制,成像系统难以无失真地获取原始场景中的信息,产生变形、模糊、降采样和噪声等问题,针对上述情况下降质图像的复原问题,提出了适用于低分辨率,低先验知识情况下的人脸复原方法,通过基于图像相似性的期望块log相似性EPLL (expected patch log likelihood)框架来构建人脸复原效果的失真函数,利用生成对抗网络的图像补全式生成过程来复原图像。所提算法在加噪率50%以及更高情况下可以保持较好的人脸图像轮廓与视觉特点,在复原加噪20%的降质图像时,相比传统的基于图像块相似性的算法,本文算法复原结果的统计特征峰值信噪比PSNR (peak signal-noise ratio)与结构相似度SSIM (structural similarity)值具有明显优势。

Human face restoration based on natural patch likehood and generative image inpainting

In the process of image acquisition, due to the limitation of imaging distance, imaging device resolution and other factors, it is difficult for the imaging devices to acquire all the significant information in the original scene without distortion, resulting in deformation, blur, down sampling results with noise. This paper proposes a resolution aiming at low-resolution and low prior knowledge of the human faces restoration. We take the image similarity based Expected Patch Log Likelihood (EPLL) framework to construct the distortion function of the face restoration effect, and use generative adversarial networks with completion function to restore the image. The proposed algorithm can maintain good contour and visual features of face images with 50% added noise and higher. Compared with traditional methods based on the image block similarity, this method can restore the 20% degraded image better. The statistical characteristics of the proposed algorithm have obvious advantages in PSNR and SSIM values in the experimental results.