基于稀疏表示的文本图像超分辨率重建研究

发布时间：2018-01-22 10:16

本文关键词： 稀疏表示字典优化双峰限制全局约束边缘增强　出处：《太原理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：伴随着信息化的高速发展,要求信息处理技术不断完善,在大多数数字图像应用中,图像处理和分析通常需要高分辨率图像或视频。当今,硬件成本已不是问题,但是对于大多数本就受到污损的低分辨率文本图像来说,即使硬件设备足够清晰,当中的文字也无法清晰显现,在这种情况下,文本图像超分辨率重建技术尤为重要。国内外学者对超分辨率重建进行了许多研究,他们的算法都已成功应用于自然图像,但应用于文本图像时效果不佳。文本图像是一种独特的图像,应研究适合它的具体技术。有些学者虽提出一些针对文本图像的算法,但存在两个问题,一是算法复杂度高,二是在先验信息不足的情况下重建效果不好。因此,本文以文本图像的特征为基础,针对稀疏表示的重建方法进行研究,在效率和精度两方面进行改进,具体研究工作如下:(1)研究文本图像的退化模型,分析当前的几种图像重建模型和字典训练算法,对原始稀疏表示的重建算法具体流程进行研究和分析。(2)对稀疏表示中的联合训练方法进行研究和分析,针对联合训练算法运算时间长,执行效率低的问题,提出一种优化的字典训练方法,只需学习高分辨率字典,近而由其推导得到低分辨率字典,从而缩短了运算时间,提高了算法的效率。在高分辨率字典学习阶段,使用K-SVD算法来训练字典;求解稀疏表示系数阶段,通过分析稀疏表示系数的局部模型,使用高效的特征符号方法进行求解。最后进行了实验仿真和分析,对算法的重建效果和执行时间都作了对比实验,运行时间缩短了45.7%,PSNR值和SSIM值稍高于原始的稀疏表示算法,证明算法在保证精度的同时提高了执行效率。(3)在优化的字典训练方法的基础上,对文本图像的特性进行研究,针对原始稀疏表示算法重建的图像不清晰,前景和背景区分不明显,不能清晰显现文字,边缘不连续的问题,对全局约束进行改进,引入文本图像的双峰限制特性作为正则项来约束重建高分辨率图像,并使用边缘增强算法来优化增强图像的边缘。对算法实验验证并与经典的稀疏重建方法以及当前两种文本图像重建的方法进行对比和分析。结果证明本文算法重建的图像边缘恢复得更好,文字和背景区分更明确。(4)分析碑文图像的特性,根据其特性设计图像处理流程,首先对它预处理,并使用本文改进的重建方法对预处理后的碑文进行SR重建。最后通过实验验证本文方法在碑文图像恢复中的可行性和实用性,实验证明经本文超分辨率重建后的图像,文字边缘清晰,前景背景区分明显,文字容易识别。
[Abstract]:With the rapid development of information technology, information processing technology is constantly improved. In most digital image applications, image processing and analysis usually require high-resolution image or video. Hardware costs are no longer a problem, but for most low-resolution text images that are already tainted, even if the hardware is clear enough, the text in the middle is not clear enough, in this case. Text image super-resolution reconstruction technology is particularly important. Scholars at home and abroad have done a lot of research on super-resolution reconstruction, their algorithms have been successfully applied to natural images. But the application of text image is not good. Text image is a kind of unique image, which should be studied. Some scholars put forward some algorithms for text image, but there are two problems. One is the high complexity of the algorithm, the other is that the reconstruction effect is not good in the case of lack of prior information. Therefore, based on the features of text images, this paper studies the sparse representation reconstruction method. The research work is as follows: 1) the degradation model of text image is studied, and several image reconstruction models and dictionary training algorithms are analyzed. The concrete flow of the original sparse representation reconstruction algorithm is studied and analyzed. (2) the joint training method in sparse representation is studied and analyzed, and the joint training algorithm takes a long time. In this paper, an optimized dictionary training method is proposed, which only needs to learn high-resolution dictionaries and get low-resolution dictionaries from them, thus shortening the operation time. Improve the efficiency of the algorithm. In the learning stage of high-resolution dictionary, K-SVD algorithm is used to train the dictionary; At the stage of solving sparse representation coefficient, the local model of sparse representation coefficient is analyzed, and the efficient characteristic symbol method is used to solve the problem. Finally, the experimental simulation and analysis are carried out. The reconstruction effect and execution time of the algorithm are compared, and the running time is reduced by 45.7% PSNR value and SSIM value slightly higher than the original sparse representation algorithm. It is proved that the algorithm not only ensures the accuracy but also improves the execution efficiency. (3) on the basis of the optimized dictionary training method, the characteristics of the text image are studied, and the image reconstructed by the original sparse representation algorithm is not clear. The distinction between foreground and background is not obvious, the text can not appear clearly, the edge is not continuous, the global constraint is improved, and the bimodal constraint feature of text image is introduced as the regular item to reconstruct high-resolution image. The edge enhancement algorithm is used to optimize the image edge enhancement. The algorithm is verified by experiments and compared with the classical sparse reconstruction method and the current two text image reconstruction methods. The results show that the proposed algorithm is heavy. The built image edges are restored better. The character of the inscription image is analyzed, and the image processing flow is designed according to its characteristics. The first step is to preprocess it. Finally, the feasibility and practicability of the proposed method in the restoration of inscription images are verified by experiments. The experimental results show that the text edge is clear, the foreground background is distinct, and the text is easy to be recognized after super-resolution reconstruction.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】