【作者】 王晏；

【导师】 孙怡；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2014， 博士

【摘要】 随着光学成像与图像处理的快速发展,基于X射线图像的目标自动检测越来越广泛地应用于工业、医疗以及军事安全等领域。在X射线图像成像过程中,有些目标因为较为弱小或者对X射线的吸收较弱,导致X射线图像中目标与背景的灰度对比度较低；X光的散射现象也造成图像存在较大随机噪声。上述低对比度和大噪声的现象在进行X射线图像目标自动检测时,很容易影响后续的图像分割和特征提取效果,从而产生一定的误检和漏检。因此,在X射线图像目标自动检测中,需要对图像进行平滑,尽可能去除随机噪声并保留图像的细节信息,研究有效的目标特征提取方法和目标分类方法。本文研究了X射线图像目标自动检测方法,主要创新性工作如下：(1)针对噪声消除中传统平滑方法出现边缘模糊的现象,提出了一种改进自适应mean shift算法的图像平滑方法。该方法将图像的灰度或伪彩色分布信息引入到meanshift算法的空域带宽设置以及权重函数设置中,将图像局部方差与频域结构相似度引入到值域带宽设置中,从而克服了传统平滑方法容易对图像造成模糊的现象,既去除了图像随机噪声,又保留了图像的细节信息,并使目标区域内相近的灰度值或伪彩色值分布更加均匀,保证了目标的完整性,避免后续的过分割现象。通过将该方法运用到X射线工业探伤图像以及X射线行李物品透视图像中,表明该方法针对含噪、低对比度的X射线图像可以获得较好的平滑结果。(2)针对低对比度、含噪的X射线图像中微弱的线型目标,提出了一种基于连续阈值分割的线型目标特征提取方法和基于霍夫变换的线型目标分割方法。实验结果表明,该方法提取的线型目标特征具有较好的可分性,可在低对比度、含噪X射线图像中获得正确的微弱线型目标的分割结果,克服了传统基于阈值和边缘的分割方法对该类图像分割不理想的缺点。(3)针对背景复杂、形状不规则的目标,提出了一种根据图像中不同分割区域之间的关系直接提取候选目标的方法。该方法不要求物体内部结构相对固定,不需要目标先验信息,克服了传统基于矩形框遍历方法比较耗时、不能获得目标具体轮廓的缺点。通过对不同分割参数的分割结果进行候选目标提取实验,结果表明该方法不需要过于复杂和精细的分割方法就能获得正确的候选目标。在候选目标确定的基础上,提取了目标链码角-弧弦距比特征,分析了该特征的旋转不变性。(4)针对两分类支持向量机对旋转目标分类时容易出现样本混叠的现象,分析了基于两分类支持向量机的多分类方法,并将目标链码角-弧弦距比特征的旋转特性构造在并联多分类支持向量机中。通过与两分类支持向量机相比,本文方法能在更大程度上保证目标的旋转不变性,克服两分类支持向量机中由于样本混叠导致的误检问题。(5)实现了一种基于稀疏表达的目标分类方法,细化了分类准则,并将目标的旋转特征构造在稀疏表达的样本矩阵中。该方法在一定程度上克服了支持向量机分类方法中样本特征对分类结果的影响。通过对X射线工业探伤图像以及X射线行李透视图像的检测表明,基于稀疏表达的方法中,样本特征对检测结果的影响小于支持向量机方法。更多还原

【Abstract】 X-ray is rapidly applied to medicine area after Roentgen discovered it in1895. With the development of computer science and information technique, the automatic object detection in X-ray images has been used widely in industry, medicine and military security. In X-ray imaging system, the small sizes of some objects or their weak absorption lead to the low contrast between the objects and their backgrounds. In addition, the scattering properties of X-ray cause plenty of random noise in X-ray image. It is difficult to segment the above images and extract effective features. Therefore, the effective methods of preprocessing, feature extraction and object classification are analyzed in this dissertation. The main contents are described as follows.(1)Traditional X-ray image smoothing methods may cause blurry edges. Therefore, an image smoothing method based on improved adaptive mean shift is proposed. The image color or gray information is utilized to analyze the spatial bandwidths and weights of sampling points. And the range bandwidths are obtained by combining the local variance function with the frequency structural similarity function. Consequently, the blurry edges are avoided, and the edge information is well reserved. The experimental results of X-ray industrial images and X-ray luggage images show that the noises are well reduced by the proposed method in the X-ray image with low contrast and high noises.(2) Based on image smoothing, the feature extraction of X-ray images is analyzed. To the weak line objects with low contrast and high noise, a feature extraction method based on multiple thresholds and Hough transformation is proposed. The experimental results show that the method can overcome the low contrast and high noise, and obtain the good results.(3) To complex X-ray image with irregular shape, the segmented regions are obtained firstly, and then object candidates are extracted by combining segmented regions. This method can meet the demand of different objects with different structures, requiring little priori knowledge. It also overcomes the shortcomings of time-consuming and coarse region by travelling windows. The experiments of object candidates extracted from different segmentation images show that the proposed method can obtain the correct object candidates without complex segmentation method. Additionally, the features of angle chain and ratio of arc-chord distance to chord length are extracted, and their rotational invariance is analyzed.(4) Based on the proposed features, the multi-classified Support Vector Machine (SVM) is analyzed and the proposed features are designed in parallel SVMs. By comparing two class SVMs with multiclass SVMs, the proposed SVMs ensure the rotational invariance of objects to some extent, and overcome the false detection by example overlap of two-class SVM method.(5) A classified method based on sparse representation is analyzed. The classification criteria are refined and the characteristic of object rotation is constructed in the sample matrix. The experimental results of X-ray industrial images and X-ray luggage images show that the classification of sparse representation requires less dependence on the samples than SVM method.更多还原