【作者】 刘毅；

【导师】 田生伟；

【作者基本信息】 新疆大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 水下图像增强技术在海洋生物研究、水下物体检测、水下航行器控制、海洋油气勘探等诸多领域都发挥着基础性作用。水下图像的退化问题是制约准确获取水下图像信息的关键因素,严重影响了相关领域的发展,改善水下图像质量十分迫切。造成水下图像退化的主要原因是水中的悬浮微粒和溶解质对光的吸收和散射作用而产生的光衰减。光的吸收会导致水下图像亮度降低,散射则会引起与距离相关的加性噪声,从而降低图像对比度,造成图像模糊。由于不同波长的色光在水中的衰减速率存在显著差异,随着距离增加与波长相关的衰减会导致图像产生全局色彩偏移现象。针对水下图像存在的对比度下降、细节丢失、颜色失真、全局色彩偏移等问题,本文以深度学习相关理论为基础,提出了两种基于深度学习的水下图像处理方法。具体工作如下:(1)提出了双注意力门融合网络的水下图像增强方法。该方法采用加入了空间注意力机制的U型网络处理输入图像以生成去除噪声且突出细节特征的置信度图,利用加入了通道注意力机制的卷积神经网络优化图像特征得到有效纠正了色彩偏移的特征图;最后将生成的置信度图与特征图融合得到增强图像。在合成水下图像数据集和真实水下图像数据集上的实验结果表明,与现有方法相比该方法取得了更优的水下图像增强效果。(2)提出了U型Swin Transformer网络的水下图像增强方法。该方法采用移位窗口的方式将自注意力计算局限在位移窗口中从而极大地减少了计算量;同时通过窗口移位实现了信息的跨窗口流动,便于模型捕捉全局信息。模型采用了基于Transformer实现的编码器和解码器网络,并通过跳跃连接减少了信息流动过程中重要特征丢失的问题。在合成水下图像数据集和真实水下图像数据集上实验结果表明,该方法有效地降低了水下图像的噪声干扰,较为显著地提高了图像的对比度,实现了较好的色彩还原效果,与现有方法相比具有优势。更多还原

【Abstract】 Underwater image enhancement plays a fundamental part in many fields,such as marine biological research,underwater object detection,underwater vehicle control,marine oil and gas exploration and so on.The degradation of underwater image is the key factor restricting the accurate acquisition of underwater image information,which seriously affects the development of related fields,and it is urgent to improve the quality of underwater image.The main reason for the degradation of underwater images is the light attenuation caused by the absorption and scattering of light by suspended particles and dissolved substances in water.The absorption of light causes a decrease in the brightness of underwater images,and the scattering introduces distance-dependent additive noise,which reduce the contrast of image and cause images blur.Since there are significant differences in the attenuation rates of different wavelengths of light in water,the wavelength-dependent attenuation with increasing distance can lead to the global color cast in the image.Aiming at the problems of underwater image,such as contrast reduction,detail loss,color distortion and global color cast,this paper proposes two underwater image processing methods based on depth learning theories.The specific work is as follows:(1)An underwater image enhancement method based on dual-attention gate fusion network is proposed.In this method,the U-structure network with spatial attention mechanism is used to process the input image to generate the confidence map that removes noise and highlights the detailed features.The convolution neural network with channel attention mechanism is used to optimize the image features and obtain the feature map that effectively corrects the color cast.Finally,the generated confidence map is fused with the feature map to achieve image enhancement.The experimental results on synthetic underwater image datasets and real underwater image datasets show that the proposed method achieves better underwater image enhancement effect than existing methods.(2)An underwater image enhancement method based on U-structure swin transformer is proposed.In this method,the shift window is adopted to limit the calculation of self attention to the shift window,which greatly reduces the amount of calculation;At the same time,the cross window flow of information is realized through window shift,which is convenient for the model to capture the global information.The model adopts the image encoder-decoder network based on Transformer,and reduces the loss of important features in the process of information flow through skip connections.The experimental results on synthetic underwater image datasets and real underwater image datasets show that this method effectively reduces underwater image noise,improves image contrast,and achieves better color reproduction,and has advantages over existing methods.更多还原