基於紋理先驗資訊的人臉圖像超解析度重建演算法研究

Research on super-resolution reconstruction algorithm of face image based on texture prior information

人臉圖像超解析度技術在監控安防等領域應用廣泛，但傳統的生成對抗網路模型重建人臉圖像時會出現失真，人臉特徵資訊提取與恢復中存在特徵丟失等問題。本文在生成對抗網路模型的基礎上添加紋理特徵分支和改進注意力機制，提出了一種基於混合注意力機制的人臉圖像超解析度演算法。首先，在生成器骨幹絡中添加同源殘差結構，提升淺層特徵與深層特徵的資訊融合。將空間和通道注意力複合機制融入生成器網路中分別提取其顯著性特徵，可在空間域和通道域中獲取更精准的特徵依賴關係，以融合判別資訊並增強網路的表徵能力。添加特徵紋理分支，增強重建圖像的紋理細節。在判別器中加入了譜歸一化策略，提升網路訓練的穩定性。實驗結果表明，使用改進的人臉圖像超解析度重建演算法在CelebA測試集上進行重建實驗，結果與原演算法PSNR與SSIM值相比均有所提升，且有效減少了重建人臉圖像的眼睛等重點部位的失真情況，與其他非生成對抗網路的主流演算法相比，有效提高了重建人臉圖像的真實感。

Face image super-resolution technology is widely used in monitoring and security fields, but the traditional generative adversarial network model will distort the face image reconstruction, and there are some problems such as feature loss in face feature information extraction and recovery. In this paper, a hybrid attentionmechanism based super-resolution algorithm of face image is proposed by adding texture feature branches and improving attention mechanism on the basis of generating adversarial network model. Firstly, homologous residual structure is added to the generator backbone network to improve the information fusion of shallow feature and deep feature. The spatial and channel attention complex mechanism is integrated into the generator network to extract the salient features, which can obtain more accurate feature dependencies in the spatial domain and the channel domain, so as to integrate the discriminant information and enhance the representation ability of the network. Add feature texture branches to enhance the texture details of the reconstructed image. The spectral normalization strategy is added to the discriminator to improve the stability of network training. The experimental results show that the reconstruction experiment conducted on CelebA test set by using the improved face image super-resolution reconstruction algorithm has improved the PSNR and SSIM values compared with the original algorithm, and effectively reduces the distortion of key parts of the reconstructed face image such as the eyes. Compared with other mainstream algorithms that do not generate adversative networks, the results show that the reconstruction experiment has improved the PSNR and SSIM values of the original algorithm. The realism of reconstructed face image is improved effectively.