图像处理中的Grouplet变换方法研究

发布时间：2018-05-10 08:47

本文选题：Grouplet变换 + 压缩感知　；参考：《南昌航空大学》2017年硕士论文

【摘要】：本论文在国家自然科学基金(51261024,51675258)、国家重点研发计划项目(2016YFF0203000)、江西省教育厅科学技术研究项目(GJJ150699)和广东省数字信号与图像处理技术重点实验室开放课题(2014GDDSIPL-01)共同资助下,围绕Grouplet变换为中心,针对图像重构、图像去噪、图像融合方面展开研究,结合新的压缩采样理论,提出一系列新的算法,并取得了一些创新性的成果。本文的主要内容包括以下几方面:第一章,详细论述了结合Grouplet变换与压缩感知的必要性、本课题的提出及其研究意义,系统介绍了超小波的发展及研究进展,尤其是Grouplet变换的国内外研究现状,最后给出了本文的主要内容和创新之处。第二章,结合Grouplet变换与压缩感知算法各自的优点,提出了基于Grouplet-压缩感知(Grouplet-CS)的图像重构方法。该方法的特色在于充分将Grouplet变换稀疏表示融合于压缩感知中,既最大限度的利用图像的几何特征,又消除了传统奈奎斯特采样理论造成的冗余与资源的浪费,可以进一步挖掘图像的方向、尺度等的纹理信息,使得即使很少的采样点数也可恢复出较清晰的图像质量。通过对Lena仿真与SAR图像的重构中,与小波变换压缩感知方法进行对比分析,证明了该方法一方面降低了传统方法的稀疏度和采样率,另一方面还提高了图像的重构质量。另外,还对不同的重构方法进行了对比,研究表明在相同的Grouplet稀疏表示和相同的压缩比下,ROMP算法整体优于OMP算法。第三章,引入贝叶斯压缩感知的思想,在传统贝叶斯变分算法的基础上经过改进,提出了适合二维的新的变分贝叶斯压缩感知重构算法,并结合Grouplet变换在稀疏表示方面的优势,提出了Grouplet-贝叶斯压缩感知(Grouplet-BCS)算法。提出的算法主要针对实际中图像会夹杂有噪声的情况,针对是否含噪以及含噪强度的大小选择Grouplet-BCS算法来自适应地降噪。经过Lena仿真研究,以及将其用于SAR图像的消噪中,并且与Grouplet-CS算法作比较,证明了提出的算法不仅降低了噪声对图像的污染,而且也在重构精确度方面有显著提高。第四章,论述了小波阈值消噪的特点以及存在的缺陷,针对小波阈值消噪中存在的问题,提出自适应Grouplet阈值消噪,并详细论证了其消噪原理和算法过程。在此基础上,提出了自适应Grouplet-CS算法和自适应Grouplet-BCS算法,并将其用于图像消噪中。通过仿真实验,将几类算法与传统的小波阈值消噪方法作对比,以及将其用于SAR图像中,分析各种方法的适用性。第五章,利用Grouplet变换可以消除图像的大冗余,在图像的各尺度方向、纹理上的深度挖掘的优势,结合脉冲耦合神经网络(PCNN)可以从复杂背景下获得有利信息的特点,提出了Grouplet-PCNN融合算法。通过与PCNN、NSCT-PCNN以及小波-PCNN做仿真对比,证明了经过Grouplet-PCNN融合算法得到的融合图像信息是最丰富全面的,像素也是最高的,各纹理、边缘等细节特征也是最明显的。最后通过对断口图像的工程应用,使得上述结论得到了有效验证。第六章,对各个章节作出总结,并指出仍需改进优化的地方及可以继续研究深入的发展方向。
[Abstract]:This paper is based on the National Natural Science Foundation (5126102451675258), the national key research and development project (2016YFF0203000), the science and technology research project of the Jiangxi Provincial Education Department (GJJ150699) and the open subject (2014GDDSIPL-01) of the Key Laboratory of digital signal and image processing technology in Guangdong. Such as reconstruction, image denoising, image fusion, combined with the new compression sampling theory, a series of new algorithms are proposed and some innovative achievements have been obtained. The main contents of this paper include the following aspects: Chapter 1, the necessity of combining Grouplet transform and compression perception is discussed in detail, and the proposal and research meaning of this subject are presented. It systematically introduces the development and research progress of super wavelet, especially the research status of Grouplet transform at home and abroad. Finally, it gives the main content and innovation in this paper. In the second chapter, combining the advantages of Grouplet transform and compressed sensing algorithm, a method of image reconstruction based on Grouplet- compression perception (Grouplet-CS) is proposed. The feature of the method is that the sparse representation of Grouplet transform is fully integrated into the compressed sensing, which not only makes the maximum use of the geometric features of the image, but also eliminates the redundancy and the waste of resources caused by the traditional Nyquist sampling theory, and can further excavate the texture information of the direction and scale of the image, making even a few sampling points. Compared with the wavelet transform compression sensing method in the reconstruction of Lena simulation and SAR image, it is proved that the method reduces the sparsity and sampling rate of the traditional method on the one hand, and also improves the quality of the reconstructed image on the other hand. In addition, the different reconstruction methods are also carried out. The research shows that under the same Grouplet sparse representation and the same compression ratio, the ROMP algorithm is better than the OMP algorithm. In the third chapter, the idea of Bias compression perception is introduced. On the basis of the traditional Bias variational algorithm, a new variational Bias compression perception reconstruction algorithm is proposed, which combines with the Grouplet. For the advantage of sparse representation, Grouplet- Bayesian compression perception (Grouplet-BCS) algorithm is proposed. The proposed algorithm is mainly aimed at the situation of noise in the actual image. The Grouplet-BCS algorithm is selected from adaptive noise reduction in view of whether the noise and the noise intensity of the algorithm come from the adaptive noise reduction. After the Lena simulation study, and the application of it to SAR In the image denoising, and compared with the Grouplet-CS algorithm, it is proved that the proposed algorithm not only reduces the noise pollution to the image, but also has a significant improvement in the reconstruction accuracy. Fourth chapter, the characteristics of the wavelet threshold denoising and the existing defects are discussed, and the adaptive Group is proposed for the problem of the small wave threshold de-noising. Let threshold de-noising, and demonstrates its denoising principle and algorithm process in detail. On this basis, the adaptive Grouplet-CS algorithm and adaptive Grouplet-BCS algorithm are proposed and used in image denoising. Through simulation experiments, several kinds of algorithms are compared with traditional wavelet threshold denoising method, and they are used in SAR images to analyze each other. In the fifth chapter, the fifth chapter, using the Grouplet transform to eliminate the large redundancy of the image, the advantage of the depth mining in the direction of the image and the depth of the texture, combined with the pulse coupled neural network (PCNN) can obtain the favorable information from the complex background, and put forward the Grouplet-PCNN fusion algorithm, through PCNN, NSCT-PCNN and The simulation and comparison of the wavelet -PCNN show that the fusion image information obtained through the Grouplet-PCNN fusion algorithm is the most abundant, the pixel is the highest, and the details of the texture and edge are most obvious. Finally, the conclusion is effectively verified by the engineering application of the fractured image. The sixth chapter is made to each chapter. Summarize and point out the areas that need to be improved and further research directions.

【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】