【作者】 胡勇；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2010， 博士

【摘要】 纹理分析是模式识别和计算机视觉领域中的重要研究内容,在科学研究和工程技术方面都有着非常广泛的应用前景。从上个世纪70年代起,国内外研究人员就开始了纹理的相关研究,直到现在仍然非常活跃。在模式识别和计算机视觉领域的重要期刊和会议上,每年都有大量的相关学术论文；许多国内外著名研究机构也在从事这方面的研究工作。然而,由于人类认知机制的复杂性以及实际图像纹理的丰富性,尽管国内外研究人员在纹理分析领域开展了大量的研究工作,并取得了许多研究成果,但目前已有的种种算法依然存在着许多不足。本文针对纹理分析及其应用中的重点和难点问题展开研究,对目前广泛采用的一些纹理描述方法进行了认真的研究和总结,对其理论方法和实验结果进行了深入的比较和分析,并探索了纹理分析在自然场景图像分割和路面裂缝检测中的应用。本文的主要研究内容如下：1、针对室外场景图像中的大面积阴影问题,提出了一种基于过渡区提取与Retinex模型的阴影消除算法。算法首先通过分析阴影的关键属性,采用高斯分解得到阴影候选区域的灰度分布,采用多判据方法确定阴影区域,实现了大面积阴影的有效检测。然后分析了过渡区域的性质以及现有的过渡区域提取算法,提出了一种改进的局部复杂度算法,并基于此算法提取阴影边界附近的过渡区域,由此将图像分割为阴影区域、过渡区域和正常光照区域。最后采用分区域的Retinex算法分别对阴影区域和正常光照区域提取光照,并消除光照影响,并对过渡区域进行插值,得到消除阴影后的图像。该算法对大面积阴影的检测与消除具有很好的效果,同时避免了Retinex算法在灰度变化剧烈区域产生的“光晕”现象,实验证实了算法的有效性。2、针对纹理特征提取中的尺度选择问题,分析了纹理特征的尺度变化特性以及对分类准确率的影响,阐明了导致多类纹理分类时的尺度不确定性的原因,提出了一种基于尺度无关特征的无监督纹理分类算法。算法通过提取不同尺度下的多种纹理特征,进而得到其在不同尺度下的变化趋势,并以此为特征表述图像的纹理信息。通过在纹理标准库上的实验证实,该算法有效地避免了最佳尺度的选择问题,对不同类型的多类纹理分类均取得了很好的效果,实验效果优于多种对比算法。3、针对聚类中的相似度测度选择问题,对多种距离测度进行了比较和分析,提出了一种新的测度度量方法,并将其应用于自然场景的图像分割中。在实验分析中,为了定量的评价分割效果,引入了一种量化的分割评价方案,实验数据显示,本章所提距离测度具有很好的稳定性,在不同的自然场景情况下均取得了很好的分类结果。为了适应应用系统的实时性要求,本章提出了一种基于HSV色彩空间的快速分割算法,该算法在保持较高准确率的基础上,实现了自然场景的快速分割。4、针对路面裂缝检测中的复杂纹理背景和光照不均问题,提出了一种针对路面图像的方向性纹理分析方法,并基于此方法,提出了一种基于纹理分析和支持向量机(Support Vector Machine, SVM)的裂缝自动检测算法。该算法采用了基于共生矩阵的方向性分析方法和具有旋转、平移不变性的形态描述符,用以减轻或消除强纹理背景和光照不均的影响,并提取具有鉴别意义的特征,通过训练基于核函数的SVM,实现准确检测不同路面背景下的裂缝信息。实验证实,该方法具有很好的适应性,在不同路面背景下均能准确检出裂缝。针对路面裂缝检测中的实时性问题,提出了一种基于局部二值模式(Local Binary Patterns, LBP)的预检测算法。该算法利用了LBP的快速及对光照不敏感的特性,同时考虑背景中的纹理及噪声的影响,采用两级边缘响应(强边缘和弱边缘)的思路,对LBP算子的模式分类结果重新分类。实验证实,该方法检测准确率高,适应不同纹理背景,而且速度快,可以满足工程检测的实时需要。更多还原

【Abstract】 Texture analysis is always an important issue in pattern recognition and computer vision, and also has found wide application in scientific research and industry fields. Since the 1970’s, texture analysis has been a major research area in this field. Every year, a large number of relevant research papers were published in important journals and conferences. Numbers of renowned research institutions are also engaged in this area.However, due to the complexity of human cognitive mechanisms and the richness of the actual image texture, it is still a great challenge for the existing approach. Although researchers have made lots of valuable research results, there still existing many puzzles to be resolved.This dissertation focuses on the key point in the field of texture analysis and its applications. Widely used texture analysis methods and its experiment results are reviewed, compared and analyzed. Its applications in nature scene segmentation and pavement crack detection were also explored in this dissertation. The main research can be summarized as follows.1. In order to eliminate shadows in natural scenes, an approach based on transition region detection and Retinex illumination model was proposed. The range of gray value of shadows was first extracted by Gaussian distribution. By using a multi-criterion method, shadows were effectively detected. Then, the existing transition regions detection methods were compared and analyzed; and an improved local complexity algorithm was proposed for accurate transition region extraction. Based on the improved algorithm, a Retinex based shadow elimination approach was proposed. Natural scene image was segmented to three regions, the shadows, the transition region and the sunshine region. The Retinex method was adopted in shadow and sunshine regions separately. After bilinear interpolation was used for fill up the transition region, a shadow-free image was finally obtained. The proposed method achieved good performance in shadow removal, and halo artifacts generated by Retinex algorithm in high dynamic range area was avoided. Experimental results on real scenes demonstrate the effectiveness of this algorithm.2. To solve the scale selection problem in the texture feature extraction, features’ characters at different scales and its effect in classification accuracy were first analyzed. After digging out the reasons of the uncertainty in multi-class texture classification, an unsupervised texture classification approach which based on scale-invariant features was proposed. By extracting a variety of different scales of texture features, the trends were obtained and were adopted as distinctive features for texture classification. Experimental results on standard texture database show that the proposed method achieved higher accuracy rate, and most important is the selection of the best scale was avoided.3. For the selection of the similarity measure in unsupervised texture classification, a variety of distance measurement were compared and analyzed. A novel distance measurement is proposed and then applied to nature scene segmentation. Quantitative assessment method was introduced in the experimental process. Experiment results show that the proposed distance measurement has good stability and always achieved good classification results in different nature scenes. For the real-time segmentation, a fast approach based on HSV color space was proposed. The proposed approach was implemented with high efficiency and good performance.4. In dealing with the complex texture and the uneven illumination background, a pavement crack detection approach based on orientation-selection texture analysis method and SVM was proposed. In the proposed approach, texture analysis based on GLCM and rotation-, translation-invariant shape descriptors were used against strong texture and uneven illumination background. By training SVM with kernel functions, the crack detection approach achieved high accurate rate in different pavement surface. Experiments on real pavement images confirmed the adaptability and accuracy of the proposed approach, even in strong texture and uneven illumination background.For real-time pavement crack detection and classification, a novel pre-detection algorithm based on LBP operator is proposed. Distinctive characteristics of LBP, such as fast and insensitive to illumination, were used in our approach. Complex texture and noisy background were also considered. Patterns of LBP were regrouped in order to capture different level edge response (strong edges and weak edges). Experiments on real road surface images confirmed that our approach achieved higher accuracy rate, even in strong texture background. It is more important that the approach we proposed is fast enough to meet the demand of real time detection.更多还原