【作者】 张睿；

【导师】 胡静；

【作者基本信息】 华中科技大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 如今,海面目标图像/视频在海防边防、海运贸易中都被广泛应用,但是海面的粼光、波浪和较远的观测距离等因素导致海面目标成像受干扰严重,目标像尺寸小、清晰程度低,难以看清。海面目标“看不清”的问题成为了学者们的一个研究焦点,其中利用软件方式进行超分辨率重建、增强目标信息质量成为了一个解决上述问题的重要研究方向。因此,本文对面向海面目标的超分辨率重建技术开展了深入研究。本文以无人机航拍和固定拍摄的方式实地采集了斜下视角度海面目标图像和视频数据,结合公开的下视角度舰船目标图像数据,构建了一个海面舰船目标图像超分辨率重建数据集。并在实拍数据的基础上,通过数据预处理和人工标注测试数据相结合的方法,构建了一个海面目标视频超分辨率重建数据集。上述数据集为本文的研究及实验提供了数据支持。针对海面目标图像往往存在模糊、缺少纹理细节信息等问题,本文提出了一种基于多层特征复用和对偶回归的图像超分辨率重建算法。该算法设计了一种多层特征复用模块,加强了多个层次特征信息的选取与融合;引入了对偶回归模块,通过额外的约束改善了损失函数,压缩了非线性映射空间。对比实验表明,该算法在各项指标上均取得了良好的重建效果。以4倍超分辨率重建实验为例,本文算法的平均PSNR和SSIM指标分别达到了26.852022d B和0.762827。针对浪花纹理和海面粼光会抢夺超分辨率重建模型对海面目标注意力的问题,本文提出了一种基于多注意力增强的视频超分辨率重建算法。该算法做了三个方面的改进和创新:设计了时空域注意力聚合模块,先通过空间注意力强化对海面目标的关注度,然后利用时间注意力聚合海面目标信息;改进了上采样重建模块,引入通道注意力机制,增强特征表达能力;设计了基于目标边缘信息的损失函数,在损失函数层面提高了对海面目标的注意力。对比实验结果表明,本文算法进一步抑制了浪花纹理和海面粼光对海面目标重建结果的不利影响,提高了海面目标的重建质量。视频超分辨率和图像超分辨率针对海面目标的应用不完全相同,算法可利用的信息不同,视频超分辨率能在时域获得更丰富的信息输入,图像超分辨率则对硬件实现的要求更低。本文通过对比实验验证了视频超分辨率算法较图像超分辨率算法在重建效果上的优越性。更多还原

【Abstract】 Nowadays,images and videos of sea targets are widely used in coastal defense,border defense and maritime trade.However,influence factors such as the sparkling light on the sea surface,the waves and the long observation distance cause serious interference to the imaging of sea targets.The target images are small in size and low in clarity,making it difficult to see clearly.The problem of "indiscernible" sea targets has become a research focus of scholars.Using software to perform super-resolution reconstruction and enhance the quality of target information has become a significant research direction to tackle the above problems.Therefore,this thesis conducts an in-depth research on the super-resolution reconstruction technology for sea targets.In this thesis,the image and video data of sea targets from oblique downward viewing angles are collected by means of UAV aerial photography and fixed shooting on the spot.Combined with the public image data of ship targets from downward viewing angles,a super-resolution reconstruction data of sea ship target images is constructed.set.On the basis of the collected data,a sea target video super-resolution reconstruction dataset is constructed by combining data preprocessing and manual labeling test data.The above datasets provide data support for the research and experiments in this thesis.Aiming at the problems that sea objects are often blurred and lack of texture details,this thesis proposes an image super-resolution reconstruction algorithm based on multi-layer feature reuse and dual regression.A multi-layer feature reuse module is designed to strengthen the selection and fusion of multi-layer feature information.And a dual regression module is introduced to improve the loss function through additional constraints and compresses the nonlinear mapping space.Contrast experiments show that the proposed algorithm has achieved good reconstruction results in various indicators.Taking the 4×super-resolution reconstruction experiment as an example,the average PSNR and SSIM of the algorithm in this thesis respectively reach 26.852022 dB and 0.762827.Aiming at the problem that the wave texture and sea sparkling light will grab the attention of the super-resolution reconstruction model to sea targets,this thesis proposes a video super-resolution reconstruction algorithm based on multi-attention enhancement.A spatiotemporal attention aggregation module is designed to strengthen the attention to sea targets through spatial attention and aggregate sea target information through temporal attention.The upsampling reconstruction module is improved by introducing the channel attention mechanism to enhance the feature expression ability.A loss function based on the target edge information is designed to enhance the attention to sea targets at the loss function level.The results of contrast experiments show that the proposed algorithm further suppresses the adverse effects of wave texture and sparkling light on the reconstruction results of the sea targets,and improves the reconstruction quality of the sea targets.The applications of video super-resolution and image super-resolution for sea targets are not exactly the same,and the information available to the algorithms is different.Video super-resolution can obtain richer information input in the time domain,while image superresolution requires less hardware implementation.The superiority of the video superresolution algorithm in the reconstruction effect compared with the image super-resolution algorithm has been verified by comparative experiments in this thesis.更多还原