【作者】 张德国；

【导师】 梁继民；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2006， 硕士

【摘要】 基于图像处理的故障诊断方法已经成为无损故障诊断的主要技术手段。和其他图像处理应用类似，阈值分割方法被广泛用于故障检测和定位，但是由于故障和成像模式的多样性，要设计一种普适的阈值分割算法是非常困难的。本文提出了一种是基于多分类器融合的阈值分割方法，该方法采用模糊积分将多种阈值分割算法的结果进行融合，其区别于现有分类器融合算法之处在于，融合过程不仅取决于各个分类器(分割算法)的判决输出，而且与各个分类器的判决能力有关。各个分类器的判决能力用模糊测度表示，它可以解释为单个分类器判决对最终融合判决的重要程度，通过融合来自多个信息源的特征，达到主观期望和客观评价间的最佳匹配。通过使用一组手工标定的测试图像进行算法评估，结果表明，本文提出的融合算法性能不仅优于单个阈值分割算法，而且优于基于多数表决和算术平均的分类器融合算法。许多景观和图像的认知与识别对于人眼来说是极其直观和简单的，但是对于计算机自动处理算法而言却难以做到。本文研究了基于侧抑制网络的生物视觉模型的图像分割方法。图像中的像素点不仅被其本身的亮度信息所兴奋，同时也被其周围像素点的亮度信息所抑制；它不仅接受其他像素点的抑制，同时也抑制其他的像素点。侧抑制网络模型对图像中的每个像素点作兴奋与抑制的加权和，然后对图像进行滤波，得到侧抑制图像。由于图像内容的多样性，侧抑制网络模型的加权参数在图像分割过程中起着至关重要的作用。本文提出了一种基于进化计算的模型参数优化算法，综合图像直方图形状、聚类和熵等信息，设计了一种多数性混合目标函数，以实现图像的最优分割。实验结果表明了这种基于生物视觉计算模型的图像分割算法的有效性。更多还原

【Abstract】 Fault detection and diagnosis by imaging technology has become a common approach in non-destructive testing applications. As in many other image processing applications, gray level thresholding methods are widely used to locate the defect position. Because of the variety of defect signatures and that of imaging models, it is hard to design a universal thresholding algorithm. In this paper, a new thresholding method based on multiple classifier fusion is proposed. It uses fuzzy integral to integrate the outputs of multiple thresholding algorithms. The proposed method differs from the traditional classifier fusion methods in that its fusion decision not only depending on the individual classifier’s output, but also incorporating the uncertainty of the classifier’s ability to make decision. The classifier’s decision making ability is represented by fuzzy measurement, which can be interpreted as the importance of the individual classifier’s local decision to the final decision. The optimal correspondence between subjective expectation and objective evidence is achieved by fusing the features from multiple sources. Evaluated by a set of hand segmented non-destructive images, the averaging validation index shows that the proposed method takes advantages over individual thresholding algorithm, also over majority voting and averaging fusion methods.The cognition and recognition of many landscapes and images are extremely intuitionistic and oversimplified for human eyes, but it is still another thing for the computers. In this paper, a novel image segmentation algorithm based on the biology vision lateral inhibition model is proposed. Each pixel in the image is excited by the brightness information of itself, inhibited by that of the pixels around it, and inhibits them reversely at the same time. The lateral inhibition network makes the excitation and the inhibition of every pixel in the image multiply weights respectively and sums up them as the new intensity of that pixel. The new image is filtered and an inhibited image is obtained. Because of the variety of the image content, the weighting parameters of the lateral network play a vital role in the segmentation process. In this paper, evolutionary strategy is used to search the optimal weighting parameters of the lateral inhibition model. The optimization object function is a multiple criteria combination of the image histogram, clustering and entropy information. The experimental results show that the image segmentation method based on the biology vision model is effective and efficient.更多还原