【作者】 张永新；

【导师】 陈莉；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2014， 博士

【摘要】 多聚焦图像融合是多源图像融合领域的一个重要分支,主要用于同一光学传感器在相同成像条件下获取的聚焦目标不同的多幅图像的融合处理。由于聚焦范围有限,光学成像系统不能将焦点内外的所有目标同时清晰成像,导致图像分析时需要耗费大量时间和精力。多聚焦图像融合是一种解决光学成像系统聚焦范围局限性问题的有效方法,可以有效提高图像信息的利用率,扩大系统工作范围,增强系统可靠性,更加准确的描述场景中的目标信息。目前,该技术广泛应用于交通、医疗、物流、军事等领域。多聚焦图像像素级融合是多聚焦图像融合的基础,它获得的原始信息最多,能够提供更多的细节信息。如何准确定位并有效提取源图像中的聚焦区域是多聚焦图像像素级融合的关键。由于受图像内容复杂性影响,传统的多聚焦图像像素级融合方法很难对源图像中聚焦区域准确定位,且融合图像质量并不理想。本论文针对现有多聚焦图像像素级融合方法存在的不足,在空间域内对多聚焦图像像素级融合算法进行了深入研究。论文主要研究内容如下：1、提出了基于鲁棒主成分分析(Robust Principal Component Analysis, RPCA)与脉冲耦合神经网络(Pulse Coupled Neural Network, PCNN)的多聚焦图像融合算法。根据RPCA构建的低维线性子空间可表示高维图像数据,增强目标特征信息,对噪声具有鲁棒性的特点,将源图像在RPCA分解域的稀疏特征作为PCNN神经元的外部输入,并根据PCNN神经元的点火频率来定位源图像中的聚焦区域,增强了融合算法对噪声的鲁棒性,提高了融合图像质量。2、提出了基于RPCA与四叉树分解相结合的多聚焦图像融合算法。利用源图像稀疏矩阵的区域一致性进行块划分,有利于提高聚焦区域信息提取的完整性和准确性。此外,四叉树分解用树结构存储图像块划分结果,有利于提高源图像递归剖分的效率。该算法在自适应确定最优分块大小的基础上,利用稀疏矩阵各稀疏矩阵子块的局部特征检测源图像的聚焦区域,抑制了“块效应”对融合图像质量的影响,取得了良好的融合效果。3、提出了基于图像分解的多成分图像融合算法。利用基于(Rudin-Osher-Fatemi, ROF)模型的Split Bregman算法将源图像分解为卡通和纹理部分,用卡通成分和纹理成分中像素邻域窗口的梯度能量(Energy of image Gradient, EOG)检测聚焦区域像素,并根据融合规则对这些像素进行融合,将融合后的卡通和纹理部分合并实现图像融合。该算法提高了融合算法对源图像几何特征描述的完整性,提升了融合算法性能,改善了融合图像的视觉效果。4、提出了基于非负矩阵分解(Negative Matrix Factorization, NMF)和聚焦区域检测的多聚焦图像融合算法。利用NMF的纯加性和稀疏性,对多聚焦图像进行初始融合,利用初始融合图像与源图像间的差异图像的局部梯度特征检测聚焦区域,根据融合规则将检测到的聚焦区域进行合并得到最后的融合图像。该算法提高了聚焦区域检测准确性,改善了传统NMF融合算法所得融合图像对比度,提高了融合图像质量。最后,对本文的主要研究工作和创新点进行总结,并对未来研究方向进行了展望。更多还原

【Abstract】 Multi-focus image fusion is an important branch of multi-sensor image fusion. It is mainly employed in the fusion of different target-specific focusing on the integration of multiple images. These images are attained by the same optical sensor at the same imaging condition. Since the limited depth-of-focus of optical lenses, it is difficult to get an image that contains all the relevant targets in focus. Moreover, it is time/space and energy consuming to analysis large number of similar images. Multi-focus image fusion can solve this problem efficiently. This technology can improve the utilization of image information, extend the work’s scope of system and strengthen the reliability of system, and more precise and reliable representation of the scene can be attained. Currently, this technology is widely employed in verticals such as medical care, transportation, logistics and military.Pixel level fusion for multi-focus image is the base of image fusion, which can get more original information and provide more detail information. How to locate the focused region is the key to multi-focus image fusion and one of the difficulties. For the affection of image content, it is difficult to accurately locate and extract the focused regions in source images, which may compromise the quality of fused image. For the shortage of the existing methods, this paper studied the multi-focus image fusion methods in the spatial domain, respectively.A multi-focus image fusion method based on (robust principal component analysis) RPCA and Pulse Coupled Neural Network (PCNN) in RPCA decomposition domain is proposed. Since RPCA can represent the high dimension in a lower dimensional linear space and strengthen the target information, which is robust to noise. Thus, the sparse feature of the source images in RPCA decomposition domain is used as the external input of PCNN neuron and the sharp region are selected by the fire times of PCNN neuron, which suppresses the noise and improves the fused quality.A multi-focus image fusion algorithm based on RPCA with Quad tree (QT) decomposition is proposed. QT decomposition partitions the image based on the regional homogeneity of sparse matrix, which is useful to safeguard the integrity of regional information. QT decomposition saves the partition results with tree structure, which improves the efficiency of block division. Thus, this paper performs the QT decomposition on sparse matrix of source images in RPCA decomposition domain and determines the optimal block size based on the regional homogeneity of sparse matrix, and then, sharp regions of source images are selected based on the local feature of each sparse matrix block. Thus, the affection of blocking artifacts is suppressed and better fusion results are attained.A multi-components multi-focus image fusion algorithm based on image decomposition is proposed. Since multicomponent of image can provide more complete representation for an image, this algorithm decomposes the source images into cartoon and texture components by using Split Bregman algorithm of Rudin-Osher-Fatemi (ROF). The focused pixels are detected by the energy of image gradient (EOG) of the neighborhood of each pixel of the cartoon and texture components. The focused pixels are fused in different fusion rule, respectively. The final fused image can be attained by merging the cartoon and texture components. This algorithm makes up for the shortage of traditional fusion method and improves the performance of fusion method in representation integrity of image detail information.A multi-focus image fusion scheme based on NMF and fused regions detection is proposed. The scheme fuses the source images to construct initial fused image by using the pure additive and sparsity of the NMF. The local features of the difference between initial fused image and source images are used to detect the fused regions of source images. The focused regions are merged to construct the final fused image based on fusion rules. The scheme can efficiently improve the fusion quality and the visual effects.Finally, a summary of the research work and the contribution is presented. In addition, the future work and the target are pointed out.更多还原