图像是当今信息传递和呈现的主要载体，高分辨率图像因其具有清晰的画质和丰富的细节，为计算机视觉领域的发展提供了坚实的基础。然而，由于成像设备、外界环境等因素的干扰，实际获取到的图像较为模糊，此时可通过超分辨率重建技术来提高图像的清晰度。近年来，基于生成对抗网络的图像超分辨率重建方法恢复的图像真实自然，更符合人眼的视觉感知。但重建纹理色彩丰富的图像会出现细节模糊、纹理杂乱等现象，并且生成对抗网络训练困难，存在模型不稳定、易崩溃等情况。

针对生成对抗网络的图像超分辨率重建算法（SRGAN）恢复的图像细节不够清晰、纹理杂乱无章的问题，提出了一种基于密集残差网络的超分辨率重建算法（SRGAN-R）。首先，在网络结构中删除批量归一化层降低网络的计算复杂度，同时保证图像的色彩、亮度信息不变。然后，借鉴密集网络分层信息和残差网络快速融合底层特征与高层特征的优势，在网络中加入密集残差块结构，提升网络的表达能力。最后，在四个基准数据集和一个自采医学数据集上进行测试，结果表明，所提算法与SRGAN算法相比，加快了模型的重建速度，重建图像的平均PSNR值和SSIM值分别提升了0.488dB和0.0246，重建图像的细节更清晰、纹理更逼真。该算法虽然提高了重建图像的质量，但在实际应用场景中对算法的稳定性也有要求，因此本文进一步对算法进行优化。

针对生成对抗网络难以收敛、训练不稳定的问题，在SRGAN-R的基础上，提出了一种优化损失函数的超分辨率重建算法（SRWGAN-R）。首先，在判别网络中借鉴WGAN来优化算法，稳定网络的训练，并用优化的残差块来提取深层特征。然后，在损失函数中使用优化的感知损失、内容损失、对抗损失以及纹理损失共同监督模型训练。最后，在四个基准数据集和一个自采医学数据集上进行测试，结果表明，所提算法与SRGAN算法相比，模型收敛速度更快，更稳定。同时，重建图像的平均PSNR值和SSIM值相比SRGAN算法提高了1.088dB和0.0329，相比SRGAN-R算法提高了0.58dB和0.0187，重建图像纹理更丰富，亮度色彩信息更符合原始图像，较好地提升了重建图像的质量。

Image is the main carrier of information transmission and presentation. High-resolution image provides a solid foundation for the development of computer vision because of their clear image quality and rich details. However, due to the interference of imaging equipment, external environment, and other factors, the obtained image is relatively blurred. At this time, the definition of the image can be improved by super-resolution reconstruction technology. In recent years, the image super-resolution reconstruction method based on the generated countermeasure network restores the real and natural image, which is more in line with the visual perception of human eyes. However, when reconstructing images with rich texture and color, there will be some phenomena such as blurred details and disordered texture, and it is difficult to generate confrontation network training, and the model is unstable and easy to collapse.

In order to solve the problems of unclear image details and disordered texture restored by image super-resolution reconstruction algorithm (SRGAN) generating countermeasure network, a super-resolution reconstruction algorithm based on dense residual network (SRGAN-R) is proposed. Firstly, the batch normalization layer is deleted in the network structure to reduce the computational complexity of the network and ensure that the color and brightness information of the image remains unchanged. The advantages of hierarchical network and dense networks can be used for reference, and then the characteristics of hierarchical networks and dense networks can be quickly expressed. Finally, it is tested on four benchmark data sets and one self-collected medical data set. The results show that compared with the SRGAN algorithm, the proposed algorithm accelerates the reconstruction speed of the model, improves the average PSNR value and SSIM value of the reconstructed image by 0.488dB and 0.0246 respectively, and the details of the reconstructed image are clearer and the texture is more realistic. Although the algorithm improves the quality of the reconstructed image, it also requires the stability of the algorithm in the actual application scene. Therefore, this paper further optimizes the algorithm.

To address the problems of difficult convergence and unstable training of generated countermeasure network, a super-resolution reconstruction algorithm based on optimized loss function (SRWGAN-R) is proposed based on SRGAN-R. Firstly, the WGAN is used to optimize the algorithm in the discrimination network to stabilize the training of the network, and the optimized residual block is used to extract the deep features. Then, in the loss function, the optimized perception loss, content loss, confrontation loss, and texture loss are used to supervise the model training. Finally, four benchmark data sets and one self-collected medical data set are tested. The results show that the proposed algorithm converges faster and is more stable than the SRGAN algorithm. At the same time, the average PSNR value and SSIM value of the reconstructed image are improved by 1.088db and 0.0329 compared with the SRGAN algorithm, and 0.58dB and 0.0187 compared with the SRGAN-R algorithm. The texture of the reconstructed image is richer, the brightness and color information is more consistent with the original image, and the quality of the reconstructed image is better improved.

[1]冯博迪, 杨海涛, 李高源, 王晋宇, 张长弓. 神经网络在SAR图像目标识别中的研究综述[J]. 兵器装备工程学报, 2021, 42(10): 15-22.

[2]Pan Z, Shen H. Multispectral Image Super-Resolution via RGB Image Fusion and Radiometric Calibration[J]. IEEE Transactions on Image Processing, 2019, 28(04): 1783-1797.

[3]Wang Z, Chen J, Hoi S. Deep Learning for Image Super-resolution: A Survey[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(10): 3365-3387.

[4]刘朋伟, 高媛, 秦品乐, 殷喆, 王丽芳. 基于多感受野的生成对抗网络医学MRI影像超分辨率重建[J]. 计算机应用, 2022, 42(03): 938-945.

[5]Vella M, Zhang B, Chen W, et al. Enhanced Hyperspectral Image Super-Resolution via RGB Fusion and TV-TV Minimization[J]. 2021: 3837-3841.

[6]王殿伟, 郝元杰, 刘颖, 谢永军, 宋海军. 基于逐级反投影网络的车牌图像超分辨率重建[J]. 激光与光电子学进展, 2020, 57(16): 64-72.

[7]房垚鑫, 郭宝峰, 马超. 基于改进点扩散函数的遥感图像超分辨率重建[J]. 激光技术, 2019, 43(05): 123-128.

[8]Wang H, Zhao X, Liang S, et al. Developing Long Time Series 1-km Land Cover Maps From 5-km AVHRR Data Using a Super-Resolution Method[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 59(07): 5479-5493.

[9]郭岑, 尹增山, 高爽. 空间感知残差网络的遥感图像超分辨率重建[J]. 测绘科学, 2020, 45(05): 49-55.

[10]吴磊, 吕国强, 赵晨, 盛杰超, 冯奇斌. 基于多尺度残差网络的CT图像超分辨率重建[J]. 液晶与显示, 2019, 34(10): 1006-1012.

[11]Ren S, Li J, Guo K, Li F. Medical Video Super-Resolution Based on Asymmetric Back-Projection Network With Multilevel Error Feedback[J]. IEEE Access, 2021, 9: 17909-17920.

[12]刘颖, 朱丽, 林庆帆, 李莹华, 王富平, 卢津. 图像超分辨率技术的回顾与展望[J].计算机科学与探索, 2020, 14(02): 181-199.

[13]Goodfellow I, Pouget-Abadie J, Mirza M, et al.Generative adversarial nets[C]//Advances in Neural Information Processing Systems. 2014: 2672-2680.

[14]夏皓, 吕宏峰, 罗军, 蔡念. 图像超分辨率深度学习研究及应用进展[J]. 计算机工程与应用, 2021, 57(24): 51-60.

[15]Dong C, Loy CC, He K, Tang X. Image Super-Resolution Using Deep Convolutional Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(2): 295-307.

[16]Dong C, Loy C, Tang X. Accelerating the Super-Resolution Convolutional Neural Network[C]//European Conference on Computer Vision, Springer, 2016: 391-407.

[17]Shi W, Caballero J, Huszár F, et al. Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2016: 1874-1883.

[18]Kim J, Lee J K, Lee K M. Accurate Image Super-Resolution Using Very Deep Convolutional Networks[C]//CVPR, 2016, 1: 1646-1654

[19]Kim J, Lee J K, Lee K M. Deeply-Recursive Convolutional Network for Image Super Resolution[C]//IEEE Conference on Computer Vision and Pattern Recognition. IEEE Computer Society, 2016: 1637-1645.

[20]Ying T, Jian Y, Xiao L. Image Super-Resolution via Deep Recursive Residual Network[C]//IEEE Conference on Computer Vision and Pattern Recognition. IEEE Computer Society, 2017: 3147-3155.

[21]Mao X, Shen C, Yang Y. Image Restoration Using Convolutional Auto-encoders with Symmetric Skip Connections[C]//NIPS, 2016.

[22]Lai W, Huang J, Ahuja N, et al. Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution[J]. IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA , 2017: 5835-5843.

[23]Tong T, Gen L, Xie L, Qin G. Image Super-Resolution Using Dense Skip Connections [C]//ICCV, 2017: 4799-4807.

[24]Zhang Y, Tian Y, Kong Y, Zhong B. Residual dense network for image super-resolution [C]//The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018: 2472-2481.

[25]Muhammad H, Greg S. Deep Back-Projection Networks For Super-Resolution[C]//The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018: 1664-1673.

[26]Lim B, Son S, Kim H, Nah S, and Lee K. M. Enhanced Deep Residual Networks for Single Image Super-Resolution[C]// IEEE Conference on Computer Vision and Pattern Recognition(CVPR)Workshops, 2017: 136-144.

[27]Zhang Y, Li K, Li K,Wang L, Zhong B, Fu Y. Image Super-Resolution Using Very Deep Residual Channel Attention Networks[C]//ECCV, 2018: 286-301.

[28]Hu S, Guo D, et al. Accurate Image Super-Resolution Using Dense Connections and Dimension Reduction Network[J]. Multimedia Tools and Applications, 2020, 79(1): 427-1443.

[29]Ledig C, Lucas T, Ferenc H, Jose C, Andrew A, Alykhan T, Johannes T, Zehan W, Wenzhe S. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2017: 4681-4690.

[30]Huang Z, Liu D, Chen W, Lu P, Chen J. Adversarial Learning for Image Super Resolution Using Auxiliary Texture Feature Attributes[C]//Asia-Pacific Conference on Communications Technology and Computer Science(ACCTCS). IEEE, 2021: 132-137.

[31]Mechrez R, Talmi I, Shama F, et al. Maintaining Natural Image Statistics with the Contextual Loss[J]. Springer, Cham, 2018.

[32]毕玉萍. 基于改进的生成式对抗网络的图像超分辨率重建算法研究[D]. 青岛大学, 2020.

[33]郑喆. 基于生成对抗网络的图像超分辨率重建算法研究[D]. 浙江师范大学, 2021.

[34]李佳星, 赵勇先, 王京华. 基于深度学习的单幅图像超分辨率重建算法综述[J]. 自动化学报, 2021, 47(10): 2341-2363.

[35]He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016: 770-778.

[36]Huang G, Liu Z, Van Der Maaten L. Densely Connected Convolutional Networks[C] //Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017: 4700-4708.

[37]Arjovsky M, Chintala S, Bottou L. Wasserstein Gan[J]. Neural Information Processing System Foundation, 2017(12): 5768-5778.

[38]郭海涛, 李宸尧, 崔君, 余东行, 卢俊, 高慧. 多尺度并联卷积神经网络遥感影像油罐目标检测算法[J]. 测绘科学技术学报, 2021, 38(02): 148-154.

[39]Park S, Gach H M, Kim S, et al. Autoencoder-Inspired Convolutional Network-Based Super-Resolution Method in MRI[J]. IEEE Journal of Translational Engineering in Health and Medicine, 2021, 99(06): 1-13.

[40]王鑫, 王翠荣, 王聪, 苑迎. 双通道多感知卷积神经网络图像超分辨率重建[J]. 东北大学学报(自然科学版), 2020, 41(11): 1564-1569.

[41]方玉明, 眭相杰, 鄢杰斌, 刘学林, 黄丽萍. 无参考图像质量评价研究进展[J]. 中国图象图形学报, 2021, 26(02): 265-286.

[42]朱海琦, 李宏, 李定文, 李富. 基于生成对抗网络的单图像超分辨率重建[J]. 吉林大学学报(理学版), 2021, 59(06): 1491-1498.

[43]Li C, Wang L, Cheng S, et al. Generative Adversarial Network-Based Super-Resolution Considering Quantitative and Perceptual Quality[J]. Symmetry, 2020, 12(3):449-450.

[44]Lan R, Sun L, Liu Z. Cascading and Enhanced Residual Networks for Accurate Single Image Super-Resolution[J]. IEEE Transactions on Cybernetics, 2021, 51(01): 115-125.

[45]胡高鹏, 陈子鎏, 王晓明, 张开放, 黄增喜, 杜亚军. 基于深度残差反投影注意力网络的图像超分辨率[J]. 计算机应用研究, 2020, 38(04): 1271-1276.

[46]Ying Tai, Jian Yang, and Xiaoming Liu. Image Super-Resolution via Deep Recursive Residual Network[C].//IEEE Conference on Computer Vision and Pattern Recognition. IEEE Computer Society, 2018: 3147-3155.

[47]吴鹏, 林国强, 郭玉荣, 赵振兵. 自学习稀疏密集连接卷积神经网络图像分类方法[J]. 信号处理, 2019, 35(10): 1747-1752.

[48]杨文霞, 王萌, 张亮. 基于密集连接块U-Net的语义人脸图像修复[J]. 计算机应用, 2020, 40(12): 3651-3657.

[49]Ni T, Ding X, Wang Y, et al. Spectrum Sensing via Temporal Convolutional Network[J]. China Communications:, 2021, 18(9): 37-47.

[50]程哲, 白茜, 张浩, 王世普, 梁宇. 使用深层卷积神经网络提高Hi-C数据分辨率[J]. 计算机科学, 2020, 47(S1): 70-74.

[51]薛英杰, 韩威, 周松斌, 刘忆森. 基于生成对抗单分类网络的异常声音检测[J]. 吉林大学学报(理学版), 2021, 59(06): 1517-1524.

[52]黄淑英, 汪斌, 李红霞, 杨勇, 胡威.基于生成对抗网络的图像去雾算法[J]. 模式识别与人工智能, 2021, 34(11): 990-1003.

[53]林志鹏, 曾立波, 吴琼水. 基于生成对抗网络的宫颈细胞图像数据增强[J]. 科学技术与工程, 2020, 20(28): 11672-11677.

[54]Mirza M, Osindero S. Conditional Generative Adversarial Nets[J]. Computer Science, 2014: 2672-2680.

[55]Wang X, Yu K, Wu S, et al. Esrgan: Enhanced Super-Resolution Generative Adversarial Networks[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 1-16.

[56]Sajjadi M, Scholkopf B, Hirsch M. EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis[J]. IEEE, 2017: 4501-4510.

Madhu M, Ramanarayanan S, Ram K, et al. Reference-based Texture transfer for Single Image Super-resolution of Magnetic Resonance images[J]. 2021: 579-583.