【作者】 陈浩；

【导师】 李世银；

【作者基本信息】 中国矿业大学 ， 信息与通信工程， 2020， 硕士

【摘要】 随着现代科技的不断发展,显微设备以其超高分辨率在人类认识微观世界中发挥了巨大作用。然而,在图像采集与记录过程中,由于设备暗电流、样品物理性质以及操作人员的主观判断等影响,会使图像出现不同程度的噪声和对比度失衡状况。这些失真状况导致显微图像质量下降,而直接对降质图像进行观察研究,会使研究人员对目标的分析产生误差。因此,对显微图像进行预处理以提高图像质量是十分重要的步骤。本文以显微仪器中重要的两类设备:扫描电子显微镜（Scanning electron microscope,SEM）和X射线显微断层成像系统（X-ray microtomography,XRM）的图像为研究对象,分析影响两类图像质量的主要失真因素,研究图像的去噪与增强预处理方法,改善图像细节特征,为后续目标分割与识别提供高质量图像。本文的主要研究内容包括以下两点:一、针对XRM煤粉图像因煤中矿物含量较低且密度差小而引起的图像对比度失衡问题,提出基于人类视觉系统的多尺度高动态范围图像色调映射方法。为丰富目标图像特征细节,本方法利用引导滤波对图像进行多尺度分解,提取各图层中的结构特征;针对图像对比度失衡以及视觉效果较差的状况,本方法基于人类视觉系统的感知效果构造原始直方图,并采用单位步长减少像素计数的方式构造新直方图,最后通过新旧直方图的加权组合构造映射直方图;将多尺度分解提取的特征加入改进后的映射直方图中,在丰富图像特征的同时实现对比度增强。实验结果表明,本方法不仅有效改善了图像视觉效果,而且在客观图像质量评价方面也获得了较高肯定。二、针对SEM成像过程中图像易受扫描时间和设备电路影响产生噪声的问题,提出基于改进U-net网络的SEM图像去噪算法。考虑到原始U-net网络中修正线性单元（Rectified linear unit,ReLU）在进入负区间后输出值为零而导致神经元静默的问题,引入泄露修正线性单元（Leaky rectified linear unit,LReLU）激励函数进行修正,使得阈值为负的神经元在网络训练时得到训练和更新,同时提升网络拟合能力;原始U-net网络训练需要大量成对的含噪图像和清晰图像,然而SEM清晰图像获取难度较大,因此本方法引入L₂损失函数,基于其用相同条件期望的噪声分布替换无噪分布不改变网络学习内容的特性,在只使用噪声图像进行网络训练的条件下,保持原有网络去噪效果。经实验证明:改进后的U-net网络在SEM图像集上仅使用含噪图像进行训练,去噪效果良好,同时降低训练数据集制备难度。该论文有图23幅,表4个,参考文献98篇。更多还原

【Abstract】 With the continuous development of modern science and technology,microscopic equipment with its ultra-high resolution has played a huge role in human understanding of the microscopic world.However,in the process of image acquisition and recording,due to the influence of the dark current of the device,the physical properties of the sample,and the subjective judgment of the operator,the image will have varying degrees of noise and contrast imbalance.These distortion conditions lead to the degradation of the quality of the microscopic image,and the direct observation and study of the degraded image will cause errors in the analysis of the target.Therefore,it is very important to preprocess the microscopic image to improve the image quality.This thesis takes the two important types of equipment in microscopy instruments:Scanning electron microscope（SEM）and X-ray microtomography（XRM）images as the research object.This thesis analyzes the main distortion factors that affect the quality of the two types of images,studies the image denoising and enhancement preprocessing methods,improves the details of the image,and provides high-quality images for subsequent target segmentation and recognition.The main research contents of this thesis include the following two points:First,to solve the problem of imbalance in image contrast caused by low mineral content in coal and small density difference in XRM coal image,a multiscale high dynamic range image tone mapping method based on human visual system is proposed.In order to enrich the feature details of the target image,this method uses guided filter to decompose the image at multiple scales to extract the structural features in each layer;for the imbalance of the image contrast and the poor visual effect,this paper constructs the original histogram based on the perception effect of the human visual system,and use the unit step to reduce the pixel count to construct a new histogram.Finally,a weighted combination of the old and new histograms is used to construct a mapping histogram;the features extracted by multiscale decomposition are added to the improved mapping histogram to enrich the image features while achieving contrast enhancement.Experimental results show that this method not only effectively improves the visual effect of the image,but also obtains high affirmation in objective image quality evaluation.Second,in view of the problem that the image is susceptible to noise caused by the scanning time and the device circuit during the SEM imaging process,an SEM image denoising algorithm based on an improved U-net network is proposed.Considering the problem that the output value of rectified linear unit（ReLU）in the original U-net network is 0 after entering the negative range,which leads to the neuron silence,the leaky rectified linear unit（LReLU）is introduced to modify,so that the neuron with negative threshold can be trained and updated during the network training,and the network fitting ability can be improved at the same time;The original u-net network training needs a large number of pairs of noisy images and clear images,but the SEM clear image acquisition is difficult,so this method introduces L₂ loss function to maintain the original network under the condition of using only noise images for network training denoising effect.This loss function can replace the noiseless distribution with the noise distribution by the same conditional expectation without changing the network learning content.It is proved by experiments that the improved U-net network uses only noisy images for training on the SEM image set,which has a good denoising effect and reduces the difficulty of making the training data set.The thesis has 23 figures,4 tables and 98 references.更多还原