【作者】 杨雷；

【导师】 李志刚； 贺占权；

【作者基本信息】 哈尔滨工程大学 ， 电子与通信工程（专业学位）， 2022， 硕士

【摘要】 图像是人类获取信息最重要的方式之一,如何获得高质量图像成为日益紧迫的问题。在成像硬件限制和光线等客观因素干扰下,由成像设备直接获取的图像的质量往往较差,缺失纹理信息。通过提升成像设备的硬件指标以获取高质量图像,难度过大且成本较高,所以低成本高效率的图像超分辨率重建技术应运而生。该技术应用广泛,目前在医学、公安、航空和军事等领域有普遍研究和应用。本文在现有的图像超分辨算法研究基础上,鉴于传统图像超分辨率算法存在忽略图像高频信息和图像重建结果不稳定等性能瓶颈,决定研究基于学习的重建算法,主要研究生成对抗网络在图像超分辨领域的应用,然后针对该算法在图像超分辨重建过程中忽略中层特征和特征提取尺度单一等不足,对其进行相关的理论研究及算法改进,提出两种图像超分辨重建算法。本文的主要研究工作可以归纳为以下两个方面:(1)本文提出了一种基于通道注意力机制和特征融合的生成对抗网络图像超分辨重建算法CABFF-SRGAN。首先在SRGAN算法的生成模型中引入通道注意力机制,给予高频信息丰富的特征通道更多权重,并结合二值化特征融合思想,融合各级基本模块的输出特征,加强算法的图像超分辨重建能力,其次去掉了基本模块中的批标准化层,防止在图像重建过程中引入伪影。在不降低判别模型分类性能情况下,以卷积层和全局平均池化层替换判别模型中的全连接层,大大减少判别模型参数。(2)本文提出了一种基于卷积核选择模块的生成对抗网络图像超分辨重建算法SKCB-SRGAN。借鉴Inception模块多尺度特征提取思想,提出基于自适应权重的卷积核选择模块,并结合通道注意力模块构建基本模块组,以此改进SRGAN算法的生成模型,提高算法对复杂结构输入图像的重建性能。在判别模型中,使用计算量较小的卷积层和全局平均池化层代替全连接层,以实现真假图像分类。为验证本文提出的CABFF-SRGAN和SKCB-SRGAN图像超分辨率重建算法的优越性,在Set5、Set14和BSDS100三个测试集下,与Bicubic、SRCNN、VDSR和SRGAN等四种常用图像超分辨算法进行对比实验,并使用峰值信噪比PSNR和结构相似性SSIM两种常用客观评价指标评估算法性能。由实验结果分析可知,本文算法获得较高的PSNR和SSIM值,重建图像具有更丰富的纹理细节信息,视觉效果更好。更多还原

【Abstract】 Images are one of the most important ways for humans to obtain information,and how to obtain high-quality images has become an increasingly urgent problem.Due to the limitations of imaging hardware and the interference of objective factors such as light,the quality of images directly obtained by imaging devices is often poor and lacks texture information.It is difficult and costly to obtain high-quality images by improving the hardware indicators of imaging equipment,so low-cost and high-efficiency image super-resolution reconstruction technology emerges as the times require.The technology is widely used,and is currently widely researched and applied in the fields of medicine,public security,aviation and military.Based on the existing research on image super-resolution algorithms,in view of the performance bottlenecks of traditional image super-resolution algorithms such as ignoring high-frequency information of images and unstable image reconstruction results,this paper decides to study a learning-based reconstruction algorithm,mainly researching generative adversarial network applications in the field of image super-resolution,and then in view of the shortcomings of this algorithm in the process of image super-resolution reconstruction,such as ignoring mid-level features and single feature extraction scale,related theoretical research and algorithm improvement are carried out,and two kinds of image super-resolution reconstruction algorithms are proposed.The main research work of this paper can be summarized into the following two aspects:(1)This paper proposes a image super-resolution reconstruction algorithm named CABFF-SRGAN based on channel attention mechanism and feature fusion.Firstly,the channel attention mechanism is introduced into the generative model of the SRGAN algorithm to give more weight to the feature channels with rich high-frequency information,combined with the idea of binary feature fusion,the output features of the basic modules at all levels are integrated to enhance the image super-resolution reconstruction capability of the algorithm.Secondly,the batch normalization layer in the base module is removed to prevent the introduction of artifacts during image reconstruction.Without reducing the classification performance of the discriminant model,the fully connected layer in the discriminant model is replaced by a convolutional layer and a global average pooling layer,which greatly reduces the parameters of the discriminant model.(2)This paper proposes a image super-resolution reconstruction algorithm named SKCB-SRGAN based on selective kernel convolution block.Based on the idea of multi-scale feature extraction of the Inception module,a convolution kernel selection module based on adaptive weight is proposed,and the basic module group is constructed by combining with the channel attention module.In this way,the generation model of SRGAN algorithm is improved and the reconstruction capability of the algorithm for complex structure input is improved.In the discriminative model,the fully connected layers are replaced with less computationally intensive convolutional layers and global average pooling layers to achieve real and fake image classification.In order to verify the superiority of the CABFF-SRGAN and SKCB-SRGAN image super-resolution reconstruction algorithms proposed in this paper,under the three test sets of Set5,Set14 and BSDS100,them are compared with the four commonly used image super-resolution algorithms such as Bicubic,SRCNN,VDSR and SRGAN.Comparative experiments are carried out,and the performance of the algorithm is evaluated using two common objective evaluation indicators: peak signal-to-noise ratio(PSNR)and structural similarity(SSIM).It can be seen from the analysis of the experimental results that the algorithm in this paper obtains higher PSNR and SSIM values,the reconstructed image has richer texture detail information,and the visual effect is better.更多还原