【作者】 周芳；

【导师】 蒋建国； 王培珍；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2011， 博士

【摘要】 焦炭是高炉炼铁的重要燃料和原料。焦炭的微观组织结构与其质量密切相关，其中的显微光学组织结构是焦炭在高炉中劣化的一个重要因素，它在很大程度上决定了焦炭的反应性和反应后强度。此外，配煤比相近的煤样，对应焦炭的显微光学组织结构十分相似，且配煤比的微小变化也会在光学组织结构中体现。因此，对焦炭显微结构中光学组织的分类识别研究，不仅可以加深对焦炭性质及其劣化行为的认识，而且对评价焦炭质量和指导炼焦配煤都具有十分重要的意义。目前国内外对焦炭显微光学组织的分析主要还是采用人工对照标准图谱进行数点统计方法或半自动检测方法。虽然近年来，数字图像处理与分析技术引起了广泛兴趣，但研究成果主要集中在焦炭的气孔参数测定上，对显微光学组织的自动识别研究进展比较缓慢，尤其是对各向异性光学组织中各小类成功识别的研究鲜见报道。基于上述背景，本文利用现代信号处理技术，通过对焦炭显微光学组织自动识别的关键技术和难点问题分析，从焦炭显微图像的空间分辨率增强、不同光学组织成分区域分割以及光学组织的自动分类识别三个方面展开研究与探讨，力图给出一些理论指导与参考方法，从而在一定程度上为炼焦配煤中提高生产效率和节约能源、降低成本提供一点有益帮助。本文主要研究工作与成果总结如下：（1）针对焦炭光学组织成分多样性和复杂性，提出一种基于两步学习的单帧焦炭显微图像分辨率增强算法。在构建两个理想图像对样本库基础上，通过动态改变训练样本和样本协方差矩阵特征向量，改进传统主分量学习算法，获取能够保持全局特征的高分辨率估计图像。通过引入“流形学习”概念，提出基于残差块的加权近邻线性嵌入流形学习算法，实现无特征量提取的细节预测补偿。最后经合成获得分辨率增强图像。实验结果表明，算法有效克服了学习方法中常见的测试样本“新数据”和图像对共生模型先验估计不足等问题。（2）在凸集优化原理基础上，通过对小波域凸集投影的可行性与优势分析，提出一种针对焦炭显微视频序列的图像分辨率增强算法。分别在小波域中构建帧间和帧内两个不同的凸集和投影算子，充分提取出隐含在相邻低分辨率图像中的细节信息，并在空域最大后验概率框架下设计一个简单的预处理共轭梯度估计器，预测观测模型中相邻因子的搜索方向与步长，约束凸集投影解的可行域，保证快速获取图像重建唯一最优解或近似最优解。实验结果验证了算法的有效性和鲁棒性。（3）针对一幅焦炭显微图像中可能包含多种不同光学组织成分，且相似成分和不同成分区域相互粘连、边界模糊不连续等特点，提出一种基于边缘置信度的改进均值偏移聚类分割算法。根据图像梯度信息设计像素的边缘置信度函数，并在传统均值偏移算法中引入权值参量，减少了运算迭代次数，提高了模式点的检测精度。通过修正空间域和色彩域特征的聚类条件，改善初次聚类结果，最终实现了光学组织区域的有效分割。（4）由于传统空域和频域方法对焦炭光学组织的分类识别效果都不太理想，提出一种基于小波轮廓波变换和局部二进制模式的多特征融合自动识别算法，实现了焦炭光学组织的较完备数学模型描述。算法首先通过对轮廓波变换的频谱混叠现象分析，提出一种小波轮廓波变换，完成图像多尺度多方向分解。并根据分解系数的边缘分布特性，提取出子频带四个统计特征参量。其次，提出改进的均匀局部二进制模式编码，分析选取出编码直方图为空域纹理特征参量。最后，设计完整的多特征融合方案，运用最近邻分类器完成焦炭光学组织的自动分类。实验结果显示，算法识别率可达90％以上，并具有良好抗干扰性能。（5）为进一步解决焦炭光学组织多样性和特征复杂性，提高自动识别精度，充分利用图像冗余信息，提出一种基于最优轮廓波包的焦炭光学组织自动识别算法。首先采用非抽样小波和非抽样方向滤波器组构建冗余轮廓波包变换，完成焦炭显微图像多尺度多方向分解。然后引入2-方向2-维主分量分析和虚拟样本概念，提出一种自适应加权的最优轮廓波包基选择算法。最后通过设计相似性判定准则，实现焦炭光学组织的自动识别。对比实验结果验证了算法的有效性、抗干扰性和鲁棒性。更多还原

【Abstract】 Coke is the important fuel and raw materials for blast furnace ironmaking. The microstructureof coke is closely related to its quality, and the microscopic optical texture structure is an importantfactor of coke’s degradation in blast furnace, which determines the coke reactivity and post reactionstrength to a great extent. In addition, when coal samples have approximate blending proportion, thecorresponding coke microscopic optical texture structure is also very similar. Moreover, the tinychanges of blending proportion will also be reflected from coke optical texture structure. Therefore,the research on the classification and identification of coke optical texture will help us furtherunderstand the nature of coke and its degradation behavior, and it also has very importantsignificance for evaluation of coke quality and guidance of coking blending.At present, however, the traditional analysis methods on coke microscopic optical texture aremainly rely on artificial numbered statistics according to the standard maps or semi-automaticdetection. Although digital image processing and analyzing technologies have attracted considerableinterest in recent years, the majority of the research results focus on the determination of cokestomatal parameters and the automatic identification research on coke microscopic optical texturedevelop slowly. Particularly, the achievements about successful identification to the small classes ofcoke anisotropic optical texture are rarely few.Based on the above background, through the key technique and difficult problem analysis forautomatic identification of coke optitcal texture, we studied and discussed the following threeaspects in this work: the first one was the spatial resolution enhancement of coke micrographs; andthe second one was the segmentation of different optical texture regions; and the last one was theautomatic classification and identification for different optical textures. The goal is to provide sometheoretical guidance and reference methods, which may provide some useful helps in a certain extentto improve the efficiency, save the energy and reduce the costs in metallurgy coke production.The main research contents and innovative contributions of this dissertation are as follows:(1) Because coke microscopic optical textures have the nature of diversity and complexity, anew two-step learning scheme was proposed in this work. Firstly, two ideal sample libraries wereconstructed with high-and low-resolution coke micrographs. The high resolution estimated imageof any a test sample would be able to preserve standard global features through improved principalcomponent analysis (IPCA) learning algorithm with dynamic training samples and eigenvectors ofcovariance matrices. Then, the concept of “manifold learning” was introduced, and the detailed localinformation of the test sample was obtained by an overlapped patch-based residue prediction usingweighted neighbor linear embedding (WNLE) manifold learning algorithm. Furthermore, theresolution enhancement image was achieved after synthesis. Numerical experimental results demonstrated that the proposed algorithm effectively overcame some common problems intraditional learning methods such as test sample "new data" and image prior estimate of symbioticmodel et al.(2) Based on the principle of convex set optimization, through the feasibility and superiorityanalysis on the convex set projecting under wavelet domain, an improved reconstruction method wasproposed to enhance the coke micrograph resolution for multi-frame video sequence. Two differentconvex sets and projection operators were designed under the wavelet-domain, from the aspects ofinter-frame and intra-frame to extract the details hidden among the adjacent observed low-resolutionframes. Furthermore, a simplified spatial-domain estimator was employed by introducing thepreconditioned conjugate gradient method to forecast the search direction and the step length ofadjacent factors in prediction model. Taking advantage of the spatial estimator to put constraints onthe potential solutions of the POCS, it could get unique optimal or near optimal solution quickly.Experimental results verified the effectiveness and robustness of the proposed algorithm.(3) Since a coke micrograph may consists of two or more different optical texturecompositions, and similar or different component regions have the characteristics with fuzzy anduncertain borders, an improved mean-shift clustering segmentation algorithm was put forward basedon the edge of confidence. Firstly, the edge confidence function was designed according to the imagegradient information. Secondly, based on this function, weight parameters are introduced to thetraditional mean-shift algorithm, which leaded to reduce the iteration times and improve theaccuracy of detected modes. Thirdly, through revising the clustering conditions in both space andcolor domains, the initial clustering results were improved, and different optical texture regions wereeffectively segmented finally.(4) As the results of the traditional methods in spatial and frequency domains are both not ideal,a fusion algorithm was proposed to bridge the gap in this work based on WBCT (Wavelet-BasedContourlet Transform, WBCT) and LBP (Local Binary Pattern, LBP), which realized the completemathematical model description of coke optical texture. Firstly, through the analysis of spectrumaliasing phenomenon about contourlet transformation, a new wavelet-base contourlet transformationwas presented, which would complete the decomposition of coke micrograph for multi-scale andmulti-direction. In particular, four features vectors were extracted out in each sub-band according tothe edge distribution features of the decomposition coefficients. Then, we proposed an improveduniform local binary pattern coding algorithm, and confirmed encoded image histogram as a spatialtexture feature. Finally, according to the designed fusion scheme and similarity measure criteria, theclasses of coke optical textures could be automaticly recognized by the nearest neighbor classifier.Extensive experimental results demonstrated the effectiveness and strong anti-interference performances of our algorithm. The recognition rate was above90%.(5) To solve the diversity and complexity problems of coke optical textures further, andimprove identification accuracy, a novel automatic recognition algorithm was developed based onoptimal contourlet packet transformation, which made full use of the image redundant information.Firstly, coke micrograph was decomposed for multi-scale and multi-direction by a nonsubsampledwavelet transformation (NWT) and a nonsubsampled directional filter banks (NSDFB). In addition,an adaptively weighted algorithm for selecting optimal basis was proposed by introducing the2-directonal2-dimensional PCA method and the virtual sample concept. Finally, the classes of cokeoptical textures were recognized according to the designed similarity measure criteria oneigenvectors of the selected basis. Experimental results are provided to validate the effectiveness,anti-interference and robustness performances of the proposed scheme.更多还原