【作者】 唐磊；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2007， 博士

【摘要】 传统的基于人工视觉检测路面病害的方法有成本高、精确度低、危险性高等诸多的不足，已经不能适应高速公路迅速发展的要求。近年来，基于图像分析的路面病害检测系统在公路养护事业中获得了广泛的应用，但当前系统的自动检测算法不完善，后期的数据处理仍然是采用人机结合甚至完全人工的方式，导致工作量仍然过大。因此，如何设计有效的自动检测算法是当前亟待解决的热点、难点问题。本文将针对自动检测算法中存在的一些难题，从路面图像的去噪、增强、灰度校正、阴影消除、病害信息的提取以及病害信息的度量、分类和地理定位等多个方面进行研究。针对路面图像噪声严重、裂缝边缘模糊且断裂点较多等问题，本文提出了两种基于偏微分方程(PDE)的路面图像增强方法。首先提出了一种基于梯度的相干增强扩散。该方法在相干增强扩散中吸纳了前向-后向扩散模型的思想，在计算扩散强度时，针对完好路面背景、裂缝边缘和噪声边缘采取不同的扩散策略，能够消除正常路面处的噪声，同时能够增强裂缝的流式结构和锐化裂缝边缘。第二种方法是针对P-M方程、Shock滤波器和相干增强扩散等基本偏微分方程模型的不足，提出了将几种基本模型加以融合的方法，在不同的图像区域采用不同的权值分配方案。该方法同样能够在增强裂缝流式结构、锐化裂缝边缘和去噪几方面均取得良好的效果。路面图像中的裂缝信息往往较为弱小，与背景的对比度很低，难以直接检测到。基于此，本文提出了一种用于路面图像的模糊增强算法。首先对模糊隶属度函数进行了研究改进：(1)对函数中的一个重要参数—渡越点，通过采用基于邻域一致性模糊熵测度的像素分类，对每个像素点均确定一对应的最佳渡越点；(2)设计了的新的函数形式，该函数具有更合适的良好的曲线形状；(3)通过自适应调整控制参数，使渡越点的位置和函数曲线进行最佳的结合。在模糊逆映射上，采用线性逆变换函数，能够同时保持模糊映射所带来的增强效果和消除由于截断带来的灰度信息的损失，在运算效率上也得到了提高。针对光照不均的路面图像，本文通过对图像进行分块采样，并采用双线性插值拟合图像照度，再对原图像和照度图像进行差值运算，消除了图像中的不均匀光照。针对阴影图像，本文提出了一种各向异性中心／环绕Retinex算法(Anisotropic Diffusion Center／Surround Retinex，ADCSR)，该算法在模拟全局照度时采用基于PDE的各向异性扩散，并根据阴影的特点出了基于“边界性”(Edge Degree，ED)的新的扩散方案，能够很好的模拟阴影图像的全局照度，因此能够在消除阴影时取得良好的效果。在裂缝信息的提取上，本文将二维平面图像映射到三维空间曲面，通过分析三维曲面中“山谷”的曲率特征，采用基于微分几何的空间检测算子准确地提取了曲面中的“山谷”，并反映射到二维图像平面作为裂缝点。对于检测结果中可能出现的伪裂缝，通过基于方向一致性的路径增长，消除过短的“路径”，成功的消除了绝大部分的伪裂缝。根据不同类型裂缝的几何形态差异，本文通过抽取裂缝的投影向量、像素总数和密度等具有鉴别意义的裂缝模式特征，设计基于BP神经网络的模式分类器实现了对裂缝的精确分类。在裂缝病害的的度量上，本文设计了有效的算法计算网状裂缝和块状裂缝的面积，以及横向裂缝和纵向裂缝的长度和宽度。在裂缝病害的地理坐标精确定位上，本文提出了模糊自适应联邦卡尔曼滤波器(Fuzzy Adaptive Federated Kalman filter, FAFKF)对GPS／DR组合导航定位系统进行数据融合。该滤波器首先通过模糊自适应滤波控制器监控观测量的残差理论值和实际值，并通过增强它们的一致性来调整各子系统观测噪声方差阵，使之更符合真实的模型，有效提高了Kalman滤波器对模型变化的适应能力。然后通过模糊自适应信息融合控制器对各子系统可信度进行模糊评判，并根据可信度自适应计算信息分配系数来实现数据的融合。更多还原

【Abstract】 Traditional manual methods for pavement distress detection have too manyshortcomings such as time-consuming、dangerous, costly and so on, so it can’t meet theneed of pavement’s development now. Nowadays, pavement distress detection based ondigital image analysis has been developed greatly. However, the algortims for automaticare still not satisfying. This paper is devoted to reseach the automatic detection algortims,including pavement surface image enhancement、shadow removal、crack imformationabstract、carck classification、crack measurement and crack positioning.In pavement surface images, the noises often very serious and cracks are very tiny, sotwo methods on Partial Differential Equation (PDE) are proposed to enhance thepavement surface. Firstly, a gradient-based coherence enhancing diffusion that enhancesthe cracks and also eliminates the other unwanted noises is proposed. The new diffusiondistinguishes cracks from unwanted elements by gradients on the assumption that thegradients of cracks are approximately unchanged, and takes different strategies to controlcracks and other elements. The new approach also absorbs the idea of forward-backwarddiffusion to determine the strengths and directions of the diffusion process in order tosharpen the edges of cracks. As a result, both the edges and the flow-like structure ofcracks are enhanced. Secondly, a new model fused by P-M diffusion、Shock filter、coherence enhancing diffusion is proposed. On the assumption that images are not noised,3 weight functions depending on the local gradients and the degree of consistency oflocal directional structure are designed, and the 3 basic PDE models are fused togetherby the 3 weight functions. Then according to the characteristics of road surface images,the new model for not-noised images are generalized to process complex road surfaceimages by improving the basic PDE models and the weight functions.The illumination of the pavement surface image often imbalance, even there aresometimes shadows in road surface images and make it difficult to process the image.Images without shadows are processed by a illumination correction method based on thesimulation of the background of the image. As for the shadows, An Anisotropic DiffusionCenter/Surround Retinex (ADCSR) is presented to eliminate it. First anisotropic diffusionbased on PDE is introduced to ADCSR, further a new anisotropic diffusion scheme basedon "Edge Degree"(ED) is presented, which avoids the embarrassment to select different parameters such as gradient threshold. Experimental results show that shadows areeliminated successfully by ADCSR.The contrast of the pavement surface images are enhanced by fuzzy imageenhancement algorithms. Traditional fuzzy image enhancement algorithms can’t enhanceimages with changeful grey levels well, also it is difficult to decide the controlparameters, so a new fuzzy image enhancement algorithm is proposed to overcome thedrawbacks. First the crossover points for each pixel are computed adaptively based onthe local feature of the neighborhood of each pixel. Then a new fuzzy membershipfunction is proposed. The new fuzzy membership function is S-shape, and can combinedwith the crossover points perfectly by adjust the parameters. Road surface images withchangeful grey levels can obtain satisfactory enhancement effect by the new algorithm.Also the new algorithm is universal because all the parameters are computed adaptively.New algorithms for detecting and classifying、measuring road surface cracks areresearched systemically. First the 2D pavement surface images are mapped to 3D spatialsurfaces and the cracks that are difficult to describe in 2D images can be regarded as"valleys" in them. Then the "valleys" are detected by differential geometry operator andtaken as cracks in the 2D images. Further the line feature of the real cracks is analyzed,and the lengths of the cracks are obtained by path growing in the consistent direction,cracks that are not long enough are considered to be fake and eliminated. Than a patternclassifier based on BP neural network is designed to recognize different cracks accordingto the geometrical shape difference of different cracks, effective methods for measuringcracks are proposed after that.At last, a fuzzy adaptive federated Kalman filter (FAFKF) is presented for positioningcracks. First, a real-time fuzzy adaptive filter controller is used to monitor the fact valueand theoretic value of residual covariance, and adjust the covariance matrices ofobservation noises towards the real model by enhancing their consistencies. As a result, theKalman filter’s tolerance to model error is improved. Then a fuzzy adaptive data fusioncontroller is used to evaluate the reliability of each subsystem, and the informationdistribution coefficients of each subsystem are computed according to the reliability.Theoretical analysis and experimental data show that the precision and fault tolerance ofFAFKF are improved both.更多还原