基于自适应卷积与联合损失函数的人脸图像超分辨率重建  

Face Image Super-resolution Reconstruction Based on Adaptive Convolution and Joint Loss Function

作　　者：李培育 张雅丽[1] 张奕博 赵益辰 LI Pei-yu;ZHANG Ya-li;ZHANG Yi-bo;ZHAO Yi-chen(College of Information and Cyber Security,People’s Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学信息与网络安全学院,北京100038

出　　处：《科学技术与工程》2025年第6期2442-2452,共11页Science Technology and Engineering

基　　金：中国人民公安大学安全防范工程双一流创新研究专项(2023SYL08)。

摘　　要：针对当前人脸图像超分辨率重建算法模型卷积单一、感受野不足、单判别网络反馈信息不精确等问题,设计了一种基于自适应卷积与联合损失函数的算法。模型使用生成对抗网络架构,生成器方面,使用自适应卷积构造双路残差块并进一步组成高效的残差组,能自主学习在不同感受野下提取到的特征权重并补充单一支路遗漏的信息。判别器方面使用Vgg与U-net架构网络作为双判别网络,并使用双判别结果计算对抗损失,该损失与内容损失、感知损失组成联合损失函数。在celeba数据集上的实验表明,该算法与RWSA算法相比峰值信噪比(peak signal noise ratio,PSNR)值提高1.166 dB,结构相似度(structure similarity,SSIM)值提高0.037,学习感知图像块相似度(learned perceptual image patch similarity,LPIPS)值优化0.033,感知因子(perceptual index,PI)指标优化0.119,与其他多种主流算法相比在图像细节清晰度方面具有优势。Content Aiming at the problems of single convolution model,insufficient Receptive field and inaccurate feedback information of single discriminant network in current face image super-resolution reconstruction algorithm,an algorithm based on adaptive convolution and joint Loss function was designed.A generation adversarial network architecture was used by the model.On the generator side,adaptive convolution was used to construct dual path residual blocks and further form efficient residual groups.It can independently learn feature weights extracted under different receptive fields and supplement missing information from a single branch.The subpixel convolution layers were used to complete quadruple reconstruction of face images.In terms of discriminators,Vgg and U-net architecture networks were used as dual discriminant networks,and dual discriminant results were used to calculate adversarial losses.The losses,content losses,and perceptual losses form a joint loss function.Experiments on the Celeba dataset show that compared with RWSA,this algorithm improves PSNR by 1.166 dB,SSIM by 0.037,LPIPS by 0.033,and PI by 0.119,compared with other mainstream algorithms,it has advantages in image detail clarity.

关 键 词：超分辨率重建 自适应卷积 联合损失函数 生成对抗网络 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]