基于组合学习的人脸超分辨率算法

现有的基于深度学习的人脸超分辨算法大部分仅仅利用一种网络分区重建高分辨率输出图像，并未考虑人脸图像中的结构性信息，导致了在人脸的重要器官重建上缺乏足够的细节信息。针对这一问题，提出一种基于组合学习的人脸超分辨率算法。该算法独立采用不同深度学习模型的优势重建感兴趣的区域，由此在训练网络的过程中每个人脸区域的数据分布不同，不同的子网络能够获得更精确的先验信息。首先，对人脸图像采用超像素分割算法生成人脸组件部分和人脸背景图像；然后，采用人脸组件生成对抗网络（C-GAN）独立重建人脸组件图像块，并采用人脸背景重建网络生成人脸背景图像；其次，使用人脸组件融合网络将两种不同模型重建的人脸组件图像块自适应融合；最后，将生成的人脸组件图像块合并至人脸背景图像中，重建出最终的人脸图像。在FEI数据集上的实验结果表明，与人脸图像超分辨率算法通过组件生成和增强学习幻构人脸图像（LCGE）及判决性增强的生成对抗网络（EDGAN）相比，所提算法的峰值信噪比（PSNR）值分别高出1.23 dB和1.11 dB。所提算法能够采用不同深度学习模型的优势组合学习重建更精准的人脸图像，同时拓展了图像重建先验的来源。

Face hallucination algorithm via combined learning

Most of the existing deep learning based face hallucination algorithms only use a single network partition to reconstruct high-resolution output images without considering the structural information in the face images, resulting in the lack of sufficient details in the reconstruction of vital organs on the face. Therefore, a face hallucination algorithm based on combined learning was proposed to tackle this problem. In the algorithm, the regions of interest were reconstructed independently by utilizing the advantages of different deep learning models, thus the data distribution of each face region was different to each other in the process of network training, and different sub-networks were able to obtain more accurate prior information. Firstly, for the face image, the superpixel segmentation algorithm was used to generate the facial component parts and facial background image. Secondly, the facial component image patches were independently reconstructed by the Component-Generative Adversarial Network (C-GAN) and the facial background reconstruction network was used to generate the facial background image. Thirdly, the facial component fusion network was used to adaptively fuse the facial component image patches reconstructed by two different models. Finally, the generated facial component image patches were merged into the facial background image to reconstruct the final face image. The experimental results on FEI dataset show that the Peak Signal to Noise Ratio (PSNR) of the proposed algorithm is respectively 1.23 dB and 1.11 dB higher than that of the face image hallucination algorithms Learning to hallucinate face images via Component Generation and Enhancement (LCGE) and Enhanced Discriminative Generative Adversarial Network (EDGAN). The proposed algorithm can perform combined learning of the advantages of different deep learning models to learn and reconstruct more accurate face images as well as expand the sources of image reconstruction prior information.