摘要
摘要
随着科技的不断进步，人们进入了信息时代。数字图像作为一种信息传播的重要形式，其分辨率的高低以及一些浑浊的介质会影响人们获取图像中的信息。在现实世界中，有非常多的因素会影响图像的分辨率，如快门、散弹噪声、抖动、衍射极限、传感器、聚焦、颜混叠等。在物体成像中也存在着很多浑浊的介质，如水滴、颗粒、烟雾等。这些因素和介质都会导致图像的分辨率降低，以及图像中的部分信息丢失，因此，提高图像的分辨率和去除图像中的雾就显得尤为重要。当成像设备与成像环境均不够完善时，采用数学算法提高图像质量，即利用稀疏表示方法对图像进行去雾超分辨，这样做的优点是既不受硬件设备和环境条件的限制，还能使成本降低，具有广阔的应用前景。
稀疏表示理论在图像处理方面的应用备受关注，如图像去噪、人脸识别、图像超分辨率重建等。通过训练字典可以将图像补丁稀疏表示，再用最少的原子代表图像补丁，准确获取图像的纹理特征信息。本文的创新点在于利用稀疏表示方法对图像进行超分辨的同时又加入了对图像去雾的研究，得出的图片效果要好于单独去雾和单独超分辨的图像。本文主要研究内容如下：
1．利用超完备字典中适当选择的元素稀疏线性组合表示图像补丁。由此为每个输入的低分辨率补丁都到相对应的稀疏系数，通过该系数得出高分辨率输出。
2．高、低分辨率图像补丁的联合训练，可以加强高、低分辨率图像补丁对间稀疏表示的相似性。因此，低分辨率图像补丁的稀疏表示能够与高分辨率图像补丁字典一起应用，生成高分辨率的图像补丁。
3．在利用稀疏表示方法对图像进行超分辨率重建的过程中，其求解是一个不适定问题，为了进一步提高图像的重建质量，在重建过程中引入了正则化约束。
4．在广泛了解现有的图像去雾的方法后，决定将暗通道先验模型这种去雾算法和稀疏表示的图像去雾算法相结合。此方法使用暗通道先验模型来去掉图像中的雾，使用稀疏表示方法去掉图像中的细节噪声，提高图像分辨率。相对以往的暗通道先验模型来说，此方法的去雾霾效果稍好一些，保真度也稍好一些。
关键词：稀疏表示；超分辨；去雾
ABSTRACT
With the continuous progress of science and technology, people have stepped into the information age. As an important form of information transmission, the resolution of digital image and some turbid media will affect people's acquisition of information in the image. In the real world, there are many factors that affect the resolution of an image, such as shutter speed, scatter noise, shake, diffraction li
mit, sensor, focus, color aliasing and so on. There are also a lot of turbid media in object imaging, such as water droplets, particles, smoke and so on. All these factors and media will lead to the reduction of image resolution and the loss of some information in the image. Therefore, it is particularly important to improve the image resolution and remove the fog in the image. On the basis of imperfect imaging equipment and conditions, the image quality is improved by mathematical algorithm. The mathematical algorithm uses the sparse representation method to defog and super-resolution the image, which has a broad application prospect. Its advantage is not limited by hardware equipment and environmental conditions, but also can reduce the cost.
The application of sparse representation theory in image processing has attracted much attention, such as image denoising, face recognition and image super-resolution reconstruction. Through the way of training dictionary, image blocks can be sparse represented, image patches can be represented by as few atoms as possible, and texture features of image can be effectively extracted. The innovation of this paper lies in that the sparse representation method is used to carry out the super-resolution of the image, and at the same time, the research on image defogging is also added. The resulting image effect is better than that of the separate defogging and separate super-resolution images. The main research contents of this paper are as follows:
1. Image patches are represented by sparse linear combinations of elements in suitably selected over complete dictionaries. Thus, a sparse representation coefficient is found for each patch input with low resolution. Then, the sparse representation coefficient is used to generate high-resolution output.
2. Through joint training of low-resolution and high-resolution image patches for the
two dictionaries, the similarity of sparse representation between high-resolution and low-resolution image patches can be enhanced. Therefore, the sparse representation of low-resolution image patches can be applied together with the dictionary of high-resolution image patches to generate high-resolution image patches.
3. In the process of super-resolution image reconstruction using sparse representation method, its solution is an ill-posed problem. In order to further improve the image reconstruction quality, regularization constraints are introduced in the reconstruction process.
4. After widely knowing the existing image defogging methods, it is decided to combine the dark channel prior model defogging algorithm with the sparse representation image defogging algorithm. In order to improve the image resolution, it uses the dark channel prior model to remove the fog in the image, and then uses the sparse representation method to remove the detail noise in the image. Com
pared with previous dark channel prior models, this method has a good effect in defogging effect and fidelity.
Keywords: sparse representation; super-resolution; defogging
目录
摘要 .................................................................................................................................... I ABSTRACT ............................................................................................................................... II 第一章绪论 . (1)
1.1课题研究的背景及意义 (1)
1.2国内外研究现状及分析 (2)
1.2.1稀疏表示算法研究现状 (2)
1.2.2图像去雾研究现状 (2)
1.2.3图像超分辨的研究现状 (3)
1.3论文的主要研究工作及结构安排 (4)
1.3.1主要研究工作 (4)
1.3.2结构安排 (4)
第二章基础理论及应用研究 (6)
2.1图像超分辨率重建的理论基础 (6)
2.1.1图像退化模型 (6)
2.1.2超分辨重建方法概述 (7)
2.2用稀疏表示进行图像处理 (7)
2.2.1信号稀疏表示模型 (7)
2.2.2稀疏表示的优化方法 (8)
2.2.3稀疏表示的字典学习 (10)
2.3图像去雾算法 (12)
2.3.1雾霾成像模型 (12)
2.3.2非物理去雾模型 (12)
2.3.3暗通道先验去雾算法 (14)
2.4图像质量评价方法 (14)
2.5本章小结 (15)
第三章利用稀疏表示与暗通道方法对图像去雾处理 (16)
3.1利用稀疏表示方法对图像进行去噪 (17)
3.1.1算法步骤 (18)
3.2估计透射率 (18)
3.3优化透射率 (21)
3.4估计大气光 (23)
3.5实验结果及分析 (27)
3.6本章小结 (30)
第四章图像去雾后利用稀疏表示进行超分辨处理 (31)
4.1引言 (31)
4.2超分辨重构过程 (31)
4.3特征提取算法 (36)
4.4联合字典训练 (38)
4.5正则化方法自适应选择 (39)
4.6仿真实验 (41)
正则化与稀疏
4.7图像去雾后进行超分辨处理效果 (45)
4.8本章小结 (53)
第五章总结与展望 (54)
5.1全文总结 (54)
5.2展望 (54)
参考文献 (56)
致谢 (60)
在读期间发表的学术论文及研究成果 (61)

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。