【作者】 魏伟一；

【导师】 李战明；

【作者基本信息】 兰州理工大学 ， 控制理论与控制工程， 2011， 博士

【摘要】 图像分割是图像分析,模式识别和图像理解中的一项极具挑战性的基础性工作,分割结果的精确性直接影响到后续任务的有效性,因此在图像处理中具有非常重要的意义。然而,在图像的采集过程中,由于非均匀光照条件或点光源照射等因素影响,导致图像的灰度不均匀而产生背景噪声。如夜间图像和红外图像等图像的整体灰度值偏低,以及照片曝光不足或逆光导致图像中的局部灰度值偏低,导致局部信息无法分辨。这种光照不均匀在一定程度上改变了图像的原始面貌,增加了图像分割及后续图像处理的难度,因此对于非均匀光照图像的分割一般都要进行前期预处理。本文针对目前有关非均匀光照图像灰度校正与分割的基本问题,提出了一些新的参考方法和改进的应用策略,主要创新点有:1.针对非均匀光照图像的直方图不具备明显的双峰问题提出了光照鲁棒的小波域灰度拉升与快速的阈值化分割方法。首先利用小波分析技术,在图像的小波域上利用Otsu分割原理进行图像的灰度拉升和对比度增强;然后针对二维阈值分割时假设远离对角线区域的数值为0而降低了分割精度的缺陷,以及采用的Shannon熵因为具有非广延性而忽略了两个子系统之间的相互作用的影响,结合Tsallis交叉熵的非广延性特征,改进了图像的二维直方图并对其进行预先聚类减少阈值分割的数据量;最后利用粒子群优化算法,实现了最佳分割阈值的快速求解。2.对于图像灰度不均匀场的建模分析与校正中,研究了非均匀光照的表示模型,采用了Retinex模型和基函数表示非均匀光照的思想,利用曲面拟合与表示的数学方法,用正交基函数的线性组合来表示非均匀光照,从而建立了非均匀光照场的表示模型和快速的参数求解方法,实现了基于能量最小化方法的光照非均匀图像的自适应校正,可以对图像在分割前进行有效的预处理。3.在基于PCNN方法的图像分割中,首先针对非均匀光照环境的特性实现自适应确定PCNN模型的部分参数,充分考虑人眼的视觉特性,利用像素的对比度设置模型中的内部活动项连接强度??值、像素邻域信息的相关性确定连接矩阵??的值,不仅考虑了像素之间的距离因素,还结合了像素间的灰度差异,其次充分考虑了图像的空间结构性上的几何信息而采用了以图像的区域互信息熵作为PCNN分割方法迭代终止条件的判决依据,从而提出了光照鲁棒的参数自适应确定的PCNN图像分割,解决了传统PCNN方法对于非均匀光照导致的灰度不均匀图像分割效果不好的问题。4.提出了同步估计非均匀光照的FCM图像分割方法。针对图像信息的模糊性和非均匀光照的不利影响,在传统的FCM模糊聚类分割方法中,将图像非均匀光照的表示模型引入到FCM的目标函数中,利用迭代求解方法可以同步获取图像的非均匀光照估计信息和图像的模糊聚类分割结果。算法同时考虑了图像中的普通噪声和非均匀光照造成的背景噪声影响。5.在非均匀光照影响较大的彩色细胞图像的分割中,提出了光照鲁棒的基于主成分分析的分割方法。算法结合主成分分析良好的空间变换和数据降维性能对图像的RGB空间数据进行主成分分析并根据各自的贡献度选择一个或两个主成分分量进行各自的分割并合成得到最终的彩色细胞图像的分割结果。其中对第一主成分分量利用基于基函数表示的能量最小化方法进行分割并同步估计非均匀光照影响,如果同时选择了第二主成分则直接利用改进的PCNN分割方法,并将分割所获得的结果和第一分量的分割结果进行合成。算法有效实现了非均匀光照影响下的彩色细胞图像的分割。更多还原

【Abstract】 Image segmentation is one of the most challenging tasks in image analysis and pattern recognition. The importance of segmentation result has a direct impact on the effectiveness of the continuous task, and it plays a vital role in image processing.While in the processing of image acquisition, the pixel gray is inhomogeneous in image because the nonuniform illumination surroundings or under exposure lead to the whole pixel gray lower than the actual value in nightly image and infrared image. The present uneven illumination makes the image segmentation and later processing more difficult because it deteriorated the real image badly. Therefore, it is urgent to pre-process the uneven illumination image before segmentation.The thesis presents some new reference methods and some revised strategies in view of the serious shortcomings existing in present image gray correct and segmentation. The main achievements are stated as follows:1.A novel pixel gray stretch in wavelet domain and rapid image threshold segmentation algorithm is proposed in view of the case that there is no visible double peak histogram in uneven illumination image. Firstly,it stretch the image gray and enhance the image contrast by the Otsu threshold segmentation method and Retinex model in wavelet domain. Secondly,since 2-D threshold segmentation methods have some faults such as supposing partial region close to zero and utilizing shannon entropy as the optimization function which has extensive property, so it is necessary to improve the gray-neighbor gray histogram to gray-gradient histogram and to cluster the new histogram field for reducing the data size. At last, take the Tsallis cross entropy as the optimization function and calculate the optimum segmentation threshold by particle swarm optimization algorithm.2.In the modeling analysis and pixel gray correction of gray inhomogeneous in image, a intensity correction algorithm adaptively based on energy minimization is supposed by means of studying the presentation model of uneven illumination and adopting the Retinex model and the mathematical idea that the uneven illumination can be presented with the linear combination of some basis functions. The algorithm build a novel model to describe uneven illumination field in any image and their parameters are computed rapidly by energy minimization, so it can be used in pre-processing effectively before image segmentation.3.An illumination robust PCNN image segmentation algorithm is proposed. Firstly, the image intensity inhomogeneous is corrected by the uneven illumination correction method based on energy minimization and the idea that the continuous curve can be denoted by linear combination of some basis functions. Secondly, some parameters can be determined adaptively; the parameterβis set by the contrast of pixels through human visual system; the link matrix W is set by the distance and gray difference of neighbor pixels. At last, the regional mutual entropy is adopted to stop iteration because some original forms of entropy do not take into account the structural information in image.4.A robust image segmentation algorithm based on fuzzy C-means clustering is proposed which can estimate the intensity inhomogeneous simultaneously and free from the influence of uneven illumination. It introduce the model of presentation uneven intensity inhomogeneous with the linear combination of basis functions into objective function in FCM in view of the fuzziness and intensity inhomogeneous in microscopic image segmentation. As result, it obtains the segmentation results and estimates the uneven illumination fields simultaneously by means of iteration method. Meanwhile, the algorithm relieves the influence produced by both normal noise and background noise.5.An illumination robust color cell image segmentation algorithm is proposed against the influence of uneven illumination by means of the principal component analysis which can transform space and lessen dimension for multi-dimensional data effectively. Firstly, transform the color image data by principal component analysis to select first or first two components according to their own degree of contribution and segment them, then compose the segmentation results to last result. Segment the first component image by energy minimize based segmentation algorithm which can estimate the intensity inhomogeneous simultaneously,for the second component image,segment it by improved PCNN method. finally compose the two segmentation results according to their degree of contribution respectively.更多还原