融合注意力的生成式对抗网络单图像超分辨率重建  

Attention Fusion Generative Adversarial Network for Single-Image Super-Resolution Reconstruction

作　　者：彭晏飞[1] 张平甲 高艺 訾玲玲[1] Peng Yanfei;Zhang Pingjia;Gao Yi;Zi Lingling(School of Electronic and Information Engineering,Liaoning Technical University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105

出　　处：《激光与光电子学进展》2021年第20期174-183,共10页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61702241,61602226);辽宁省教育厅高等学校基本科研项目(LJ2017FBL004)。

摘　　要：基于深度学习的单图像超分辨率重建方法已经比较完善,重建图像具有较高的客观评价值或具有较好的视觉效果,但是图像感知效果和客观评价值不能均衡提升。针对这一问题,提出一种融合注意力的生成式对抗网络单图像超分辨率重建方法。首先去掉残差网络中会破坏图像原本的对比度信息、影响图像生成质量的批归一层,其次是构造注意力卷积神经网络残差块,可有效地在特征映射中进行自适应特征细化,改善重建结果在大尺度因子下缺乏高频信息和纹理细节的问题,最后构造像素损失函数,使用鲁棒性较好的Charbonnier损失函数替代均方差损失函数,用总变差正则项平滑训练结果。实验结果表明,在4倍放大因子下,与其他方法在Set5、Set14、Urban100、BSDS100测试集上进行测试比较,本文方法峰值信噪比平均值提升2.88dB,结构相似性平均值提升0.078。实验数据和效果图表明,该方法主观上具有丰富的细节,客观上具有较高的峰值信噪比值和结构相似性值,实现了视觉效果和客观评价指标值的均衡提升。Deep learning-based single-image super-resolution reconstruction method has been relatively perfect.The reconstructed image has a high objective evaluation value or a good visual effect;however,the image perception effect and objective evaluation value cannot be improved in a balanced manner.To address this problem,this paper proposes a single-image super-resolution reconstruction method based on an attention fusion generative adversarial network.In the proposed method,first,the batch layer that destroys the original image contrast information and affects the quality of image generation in the residual network is removed.Then,the residual block of the attention convolutional neural network,which can effectively perform adaptive feature refinement in the feature map,is constructed.To improve the reconstruction results that lack high-frequency information and texture details under large-scale factors,apixel-loss function is constructed to replace the mean squared error-loss function with a more robust Charbonnier loss function,and a total variation regular term is used to smooth the training results.The experimental results show that compared with other methods on the Set5,Set14,Urban100,and BSDS100 test sets under 4× magnification factor,the average peak signal-to-noise ratio and average structure similarity increased by 2.88 dB and 0.078,respectively.The experimental data and renderings demonstrate that the proposed method is subjectively rich in details,objectively has a high peak signal-to-noise ratio and structural similarity value,and achieves a balanced improvement of visual effects and objective evaluation index values.

关 键 词：图像处理 超分辨率 残差学习 卷积神经网络 注意力 生成式对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]