【作者】 辛国江；

【导师】 邹北骥；

【作者基本信息】 中南大学 ， 计算机科学与技术， 2013， 博士

【摘要】 图像分割和图像融合是图像处理的两个重要研究方向。图像分割和图像融合是目标识别、目标跟踪、目标识别等相关研究的基础工作,具有重要的研究意义,一直是人们的研究热点。由于应用的场合和分割目的的不同,图像分割形成了多种分割方法,但是没有一种通用的图像分割方法。归根结底还是计算机不能够像人一样的工作,算法的设计受到了局限。为了使计算机能够像人一样进行图像分割,计算机视觉的一些理论被逐渐引入到图像分割的研究中,形成了以脉冲耦合神经网络模型(PCNN)为代表的的基于人类视觉机理的图像分割方法。如何让这些新的算法更加智能,是研究者们关注的热点。图像融合指的是将两个或者两个以上的传感器在同一时间或者不同时间获得的关于某一场景的图像或者图像序列信息进行综合,从而生成新的关于这个场景解释的信息处理过程。如何让图像融合算法更加贴近人类视觉的工作模式,让融合图像更加符合人类的感知标准,一直是图像融合研究追求的目标。随着人类视觉机理的发展,图像融合中也逐渐引入了人类视觉的一些研究成果,这大大促进了图像融合的发展。本文以图像分割和图像融合为研究对象,引入人类视觉机理模型作为研究方法,探索图像分割和图像融合的新方法,使其更加符合人类视觉标准,取得了较好的结果。主要研究工作如下：(1)提出了一种基于脉冲耦合神经网络模型的图像分割新方法。该方法是模拟人类视觉机理的一种图像分割方法。方法将图像像素看成是大脑神经元,在外界输入的刺激下,通过循环迭代的方式发放脉冲,模拟人脑视觉皮层的工作方式。在对图像进行分割的过程中,引入了最大方差比准则对分割过程进行判定,寻找最佳分割点。通过对直方图分别为单峰、双峰和多峰的图像进行实验分析,证明了这种算法的有效性。(2)针对多图像融合问题,在标准PCNN的基础上,提出了一种新的模型——双层PCNN模型。这种模型可以更好的模拟人类视觉在处理多图像问题时的工作机理,更加符合人类视觉机理。(3)分别基于小波变换、曲波变换和轮廓波变换,同时结合PCNN模型和双层PCNN模型,提出了三种不同的图像融合方法。在利用小波变换进行图像分解的时候,利用PCNN模型和图像局部区域梯度能量来进行分解系数的选取。在利用曲波变换和轮廓波变换进行图像分解的时候,利用提出的双层PCNN模型和和局部能量匹配准则来选取系数。实验证明,本文提出的这三种图像融合方法都取得了很好的融合效果,融合结果符合人的视觉感知。相对于一些传统的图像融合方法,其对图像融合的过程更加符合人类视觉在做图像融合时的工作机理。图103幅,表4个,参考文献190篇。更多还原

【Abstract】 Image segmentation and image fusion are two important research fields of image processing. Image segmentation and image fusion are the basic work of target recognition, target tracking, target recognition. They are important for research. They have been the research focus for people.Due to the different application and different purpose, many segmentation methods are proposed. But there is not a method can solve all the segmentation problems. After all, the computer can not work as human. The algorithm design is also constrained. In order to make the computer segment the image like human, some computer vision theories are gradually introduced into the image segmentation. Many methods based on human visual mechanism are proposed, such as pulse coupled neural network model (PCNN). How to make these new algorithms smarter is the focus of the researchers.Image fusion is an information processing process, which refers to synthesize the images or sequence of images of a scene obtained by two or more sensors at the same time or different times, generating a new interpretation of the scene. It’s a destination of the image fusion that How to make the fusion algorithms closer to the work mode of the human vision, and make the fused image more in line with the standards of human perception. With the development of the human visual mechanism, some research results are introduced into the image fusion to promote the development of image fusion.In this thesis, image segmentation and image fusion are chose as the research object. To explore the new method of image segmentation and image fusion, to make it more in line with the standards of human vision, the human visual mechanism model is introduced as the research method. The main research works are as follows:(1) A new image segmentation method based on the pulse coupled neural network model is proposed. The method can simulate the human visual mechanism. In the method, the image pixel is regarded as the brain’s neurons. With the external input stimulus, the pixel generate the pulse during the the loop iteration, and then simulate the work mode of the human brain visual cortex. In the image segmentation process, the maximum variance ratio rule is introduced in the segmentation process to find the best segmentation point. The experimental analysis on the images whose histograms are unimodal, bimodal and multimodal respectively proves the validity of the proposed algorithm. And it also proves the simulation effect of human vision.(2) For the multi-images fusion, a new model named dual-layer PCNN model based on the standard PCNN model is proposed. This model can better simulate the work mechanism of human vision, and it is more in line with the human visual mechanism.(3) Three different image fusion methods based on wavelet transform, curvelet transform and contourlet transform is proposed, combined with PCNN model and dual-layer PCNN model. In the method based on the wavelet transform, when the image is decomposed by the wavelet transform, PCNN model and the energy of image local area gradient are used to select the coefficients. When the image is decomposed by curvelet transform or contourlet transform, the dual-layer PCNN model and local energy matching rule are used to select the coefficient. Experiments show that, three image fusion methods proposed in this thesis have good fusion effect, the fusion results consistent with human visual perception. Compared with the traditional image fusion methods, the proposed methods do more in line with the work mechanism of human vision in the image fusion process.更多还原