【作者】 刘晓；

【导师】 邢永康；

【作者基本信息】 重庆大学 ， 工程（计算机技术）（专业学位）， 2021， 硕士

【摘要】 图像的超分辨率重建,是指在不改变成像方法的基础上,通过图像处理技术,从算法层面提高图像分辨率,构建出更高质量图像。该技术目前已经在医疗、检测、通信、公共安全以及遥感成像等多个领域得到了广泛应用。近年来,随着机器学习,尤其是深度学习技术的快速发展,基于深度学习模型的图像超分辨率重建技术已经发展成为主流模型,并进而推动了图像超分辨率技术研究和应用的快速发展。本论文正是对该热点领域的探索。首先对经典的SRGAN(Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network)模型进行了研究和改进,通过对比实验,验证了改进的有效性。然后以道路交通标志图像的超分辨率重建为应用背景,验证了图像超分辨率应用的可行性。论文的主要工作如下:(1)通过查阅大量资料,选择以重建出图像更有真实感的SRGAN算法作为基础进行研究,之后对SRGAN的激活函数做了改进:分别用Swish和Mish函数替代SRGAN的PRelu和Leaky Relu函数,对比实验表明,改进后的模型有更低的损失函数损失值、更快的收敛速度和更高的训练稳定性。(2)针对SRGAN模型重建出来的图像峰值信噪比低和图像有伪影问题,提出了一种改进的模型--基于深度递归残差结构的生成对抗网络的模型SDSRGAN(Squeeze-and-Excitation Deep Recursive Residual Network Generative Adversarial Networks)。该模型的主要特点有:首先设计了一种新的残差块,残差块中添加了SE module来快速提取图像重要特征,提高全局感受野,并且去除了批规范化层来降低参数量和计算复杂度,同时将卷积层和激活层位置倒换以提高残差块的性能;其次把递归网络和残差网络结合,在不增加参数情况下加深网络层,提高图像的重建质量;最后将Adam优化器更改为RAdam优化器以方便模型的训练。实验结果表明,与现有的其他经典的超分辨率算法相比,SDSRGAN在客观评价指标PSNR/SSIM和主观评价中都表现出更好的性能。(3)将图像超分辨率重建技术应用于交通标志图像的重建上,用以提高辅助驾驶中的交通标志识别准确率。针对交通标志图像的特征和交通识别领域要求的实时性,设计了基于交通标志图像的算法模型TRSRGAN(Traffic Sign Super-Resolution Generative Adversarial Networks)。模型中改进了残差块,将混合深度卷积用到残差网络中。在损失函数上,用Charbonnier函数作为内容损失来提高重建图像的精度;并用预训练的VGG网络激活前的特征值计算感知损失,使得重建交通标志图像有更好的纹理色彩。最后通过自制一个交通标志数据集CQUNJ205以方便模型的训练,并用Cutblur的图像增强方法提高网络性能。实验结果表明改进的模型在参数减少的情况下,重建的交通标志图像在客观评价指标上优于其他对比算法,同时在纹理细节和色彩亮度上也有所提升。更多还原

【Abstract】 Image super-resolution reconstruction is to improve the image resolution and build a higher quality image through image processing technology without changing the imaging method.The technology has been applied in many fields such as medical treatment,detection,communication,public safety and remote sensing imaging.In recent years,with the rapid development of machine learning,especially deep learning technology,image super-resolution reconstruction technology based on deep learning model has become the mainstream model,which promotes the rapid development of research and application of image super-resolution technology.This thesis is the exploration of this hot field.Firstly,the classical SRGAN(Photo-Realistic Single Image Super-Resolution Using a General Adverse Network)model is studied and improved.Through comparative experiments,the effectiveness of the improvement is verified.Then,taking the super-resolution reconstruction of road traffic sign image as the application background,the feasibility of image super-resolution application is verified.The main work of this thesis is as follows:(1)By consulting a large number of data,SRGAN algorithm is selected as the basis for research,and then the activation function of SRGAN is improved: Swish and Mish functions are used to replace the PRelu and Leaky Relu functions of SRGAN respectively.The comparative experiments show that the improved model has lower loss value of loss function,faster convergence speed and higher training stability..(2)In order to solve the problems of low PSNR and artifacts in the image reconstructed by SRGAN model,a new model SDSRGAN(Squeeze-and-Excitation Deep Recursive Residual Network Generative Adversarial Networks)is proposed.The main features of the model are as follows: firstly,a new residual block is designed.Se module is added to the residual block to quickly extract important image features,improve the global receptive field,and remove the batch normalization layer to reduce the amount of parameters and computational complexity.At the same time,the position of convolution layer and activation layer is inverted to improve the performance of residual block.Secondly,the recursive network and residual network are combined to deepen the network layer without increasing the parameters to improve the quality of image reconstruction.Finally,Adam optimizer is changed to RAdam optimizer to facilitate model training.Experimental results show that SDSRGAN performs better than other classic super-resolution algorithms in both objective evaluation index PSNR /SSIM and subjective evaluation.(3)The image super-resolution reconstruction technology is applied to the reconstruction of traffic sign image to improve the accuracy of traffic sign recognition in assisted driving.According to the characteristics of traffic sign image and the real-time requirement of traffic recognition field,an algorithm model TRSRGAN(Traffic Sign Super-Resolution Generative Adverse Networks)based on traffic sign image is designed.The residual block is improved in the model,and the mixed depth convolution is used in the residual network.In the loss function,the Charbonnier function is used as the content loss to improve the accuracy of the reconstructed image,and the perceptual loss is calculated by the eigenvalues of the pre trained VGG network before activation,so that the reconstructed traffic sign image has better texture color.Finally,for the convenience of training,a traffic sign data set CQUNJ205 is made,and Cutblur’s image enhancement method is used to improve the network performance.The experimental results show that the traffic sign image reconstructed by the improved model is better than other contrast algorithms in the objective evaluation index,and the texture details and color brightness are also improved.更多还原