图像分割是计算机视觉的基础任务之一，基于图像分割的图像识别与检测已经在许多工业领域得到了广泛应用，但是由于实际工业场景中的图像往往存在或者标记样本获取代价高昂或者难以获取、数据量大、信号像素和噪声像素难以区分、边界像素难以分类等问题，导致基于图像分割的图像识别与检测存在精度低、运行时间长等缺点。因此真实场景下的图像分割仍然面临着严峻的挑战。

    近些年来，模糊聚类作为人工智能的重要技术之一，可以有效地处理图像分割中的一系列问题，得到越来越多国内外学者的关注。针对图像分割中边界像素易错分、真实场景中标记样本难以获取、分割精度低、运行时间长等问题展开深入研究。本文的主要创新点如下：

    (1) 针对自然图像分割问题中，边界像素类别容易错分、算法运行时间长，本文提出了一种犹豫模糊C均值算法。首先，将待处理的自然图像灰度化并转换为直觉模糊图像，模拟多专家决策过程，将每一个中心像素的邻域像素的隶属度用于构建犹豫模糊元。在决策者风险偏好（乐观者/悲观者）不同的场景下，我们基于犹豫模糊元构建距离测度方法，补齐长短不一的犹豫模糊元以便于计算距离，并将该距离引入传统的模糊聚类方法中构建犹豫模糊C均值算法的目标函数，通过拉格朗日乘子法求解得到类心和隶属度的更新公式，迭代求解得到最终图像分割结果。此外，我们还抽取犹豫模糊元的五个最重要特征，构建一种新型距离测度，并将其引入模糊聚类方法中。在公开数据集上进行精度与运行时间实验，并与模糊聚类系列方法完成对比，结果表明我们提出的方法在提高精度的同时降低了运行时间。

    (2) 针对自然图像分割问题中，当前方法对图像细节的刻画能力不足，本文提出了一种毕达哥拉斯模糊C均值算法。首先，将待处理的图像灰度化并转化为直觉模糊图像，模拟人类决策“赞成”，“反对”，“弃权”的过程，将每一个像素构建为毕达哥拉斯模糊元，在此基础上构建距离测度方法并引入模糊聚类中提出毕达哥拉斯模糊C均值方法。将其与具有代表性的无监督图像分割方法对比，结果表明我们提出的方法具有良好的细节刻画能力。

    (3) 针对天文领域中天籁望远镜获取的干涉条纹数据信号像素与噪声像素难以区分、条纹的识别长期依赖于人力效率低代价高、识别精度低的问题，本文提出了一种基于模糊聚类的层次化干涉条纹识别方法。该方法分为两个阶段，在第一阶段，应用一种模糊聚类方法对干涉条纹图像数据完成分割，实现信号像素和噪声像素的初步分离，在第二阶段，通过迭代地选取分割结果的最大连通区域并将该区域的宽度视为干涉条纹的宽度从而实现干涉条纹的位置识别。特别地，实际场景中的条纹识别由于情况复杂，常常需要运用一些预处理方法。在模拟数据和天籁望远镜获取的观测数据上完成本方法测试，与当前具有代表性的方法和经典图像分割方法完成对比实验，结果表明我们提出的方法可以自动地完成条纹位置的高精度识别。

    (4) 针对石油生产领域油井示功图识别精度较低、长期依赖传统方法和人力识别的问题，本文提出了一种融合模糊聚类的油井示功图识别方法。该方法首先应用模糊聚类对油井传感器采集到的示功图完成一次图像分割，凸显其形状特征，进一步构建神经网络提取示功图图像的深层次特征。在经过迭代训练之后，得到完整的识别网络。示功图数据集中的样本全部来自于真实生产场景，包含正常工况样本、异常工况样本以及错误样本，使得本文的方法更加贴近实际。与经典机器学习方法和专用于处理时间序列的代表性方法作对比，结果表明我们的方法可以明显提高示功图的识别精度，为石油安全生产提供了范例。

    (5) 针对岩性识别领域长期依赖野外勘探，存在低效率、代价高昂，以及偏远地区难以开展工作等问题，本文提出了一种面向岩性识别的模糊聚类方法。该方法首先将毕达哥拉斯模糊集及其距离测度引入传统的简单线性迭代聚类方法提出全新的毕达哥拉斯模糊超像素分割方法，将其用于高光谱遥感的假彩色影像得到超像素分割结果。在此基础上构建卷积神经网络，将高光谱遥感数据与超像素分割结果共同用于训练网络。在多次迭代之后，获得最终的高光谱遥感图像分割结果，并结合标准光谱完成岩性识别。与基于不同原理的聚类方法完成对比实验，结果表明我们提出的方法在真实遥感数据的应用中精度大幅提升且耗时可接受。

    综上，本文针对基于图像分割的图像检测与识别，结合模糊集理论及其扩展与距离测度方法，提出了犹豫模糊C均值算法和毕达哥拉斯模糊C均值算法，特别针对来自真实场景的图像，在经典模糊聚类的基础上提出基于模糊聚类的层次化干涉条纹识别方法，在犹豫模糊C均值算法的基础上提出融合模糊聚类的油井示功图识别方法，在毕达哥拉斯模糊C均值算法的基础上提出了面向岩性识别的模糊聚类方法，并收集真实数据集进行方法的验证。本文的研究一方面将为模糊聚类方法用于图像分割提供新思路和新方法，满足工业应用的实际需求；另一方面也将丰富人工智能领域中无监督学习的相关理论。

  Image segmentation is one of basic tasks in computer vision, image recognition and detection based on image segmentation have been widely applied in many industry fields, but owing to labeled samples expensive or difficult to obtain, massive data, difficulty to discriminate signal pixels and noise pixels, difficult to classify border pixels in real industry scenarios, which will lead to low accuracy, high running time and other problems in image recognition and detection based on image segmentation. Hence, image segmentation in real scenarios still face severe challenges.

   In recent years, fuzzy clustering, as one of the important technologies in artificial intelligence, can handle a series problems in image segmentation effectively, which attracts many civil and international researchers' attention. Aiming at border pixels easily misclassified, hard to obtain labeled samples in real scenarios, long running time and low accuracy in image segmentation, we conduct our research deeply. Main highlights are as follows:

   (1) Aiming at border pixels easily misclassified, long running time in image segmentation, this thesis proposes hesitant fuzzy C-means algorithm. Firstly, we transfer natural image into gray level value and obtain intuitionistic fuzzy image, simulate multiple experts decision making process, construct hesitant fuzzy elements through merging membership degree of neighbor pixels of every central pixel. Under different decision makers' risk preference scenarios (optimists and pessimists), we construct distance measure method based on hesitant fuzzy elements, we add minimal or maximum value to complement hesitant fuzzy elements with different length so that computing distances, and introduce these distances into traditional fuzzy clustering to establish objective function of hesitant fuzzy C-means algorithm and compute membership degree and center updating solutions through Lagrangian multiplier and obtain final segmentation results iteratively. In addition, we extract the five most important indexes to construct a novel distance measure and introduce it into fuzzy clustering methods. We make comparison experiments between multiple fuzzy clustering series methods in open datasets to test accuracy and running time, the results show that our method improves image segmentation accuracy meanwhile reduces running time.

   (2) Aiming at limited ability to depict image details of existing methods in image segmentation, this thesis proposes Pythagorean fuzzy C-means algorithm. Firstly, we transfer image into gray level image and obtain intuitionistic fuzzy image, simulating people decision making “positive-votes”, “negative-votes”, “neutral-votes” to construct Pythagorean fuzzy elements for every pixel. Based on these, we construct distance measure methods and introduce fuzzy clustering to propose Pythagorean fuzzy C-means algorithm. We make comparison experiments with representative unsupervised segmentation methods and the results show our method has good details depiction ability.

   (3) Aiming at difficulty to discriminate signal pixels and noise pixels of interferometric fringes images collected from Tianlai project in astronomy, fringes detection relying on human labor long time, low identification accuracy, this thesis proposes hierarchical interferometric fringes detection method based fuzzy clustering. This method contains two stage. In first stage, we apply fuzzy clustering method to complete image segmentation for interferometric fringes image to realize signal pixels and noise pixels preliminary separation. In second stage, we can realize location recognition of interferometric fringes through iteratively selecting maximum connection region of segmentation results and regard the width of regions as the location of fringes. Specially, fringes detection in real scenario is complex, we usually use some preprocessing methods. We make comparison experiments in simulated data and Tianlai telescope data with representative methods and classic image segmentation methods. The results show our proposed method can complete fringes location recognition with high accuracy and automatically.

   (4) Aiming at low accuracy of dynamometer cards identification, relying on traditional methods and human labor long time in oil production field, this thesis proposes combining fuzzy clustering oil well dynamometer cards identification method. This method firstly applies fuzzy clustering to segment dynamometer cards collected by sensors from oil well to protrude shape features, further constructs neural network to extract deep features of dynamometers cards. After iterative training, we can obtain complete identification network. All these dynamometer cards are from real production scenarios, containing normal working condition, abnormal working conditions and error samples, which makes this method closer to reality. We make comparison experiments with classic machine learning methods and representative methods special for time series identification, the results show our method can improve accuracy of dynamometer cards identification obviously, provide example for oil safety production.

   (5) Aiming at dependence on field exploration in long term, low efficiency and high costs, hard to conduct regional area research in lithology recognition field, this thesis proposes a novel fuzzy clustering method oriented on lithology recognition. We firstly introduce Pythagorean fuzzy sets and its distance measures into classic simple linear iterative clustering (SLIC) to propose novel Pythagorean fuzzy SLIC super pixels segmentation method, we apply it in high spectral pseudo color image and obtain superpixels segmentation results. We construct convolutional neural networks based on this method, and cooperate with high spectral remote sensing data and superpixels results to complete training process. After multiple iteration, we obtain final high spectral remote sensing images segmentation results. With the help of standard spectral, we complete lithology recognition. Comparing with clustering methods based on different principles, we complete comparison experiments, the results show our proposed method has significant accuracy improvement in real remote sensing data application and running time is acceptable.

   In summary, aiming at image recognition and detection based on image segmentation, combining fuzzy sets theory and its extension and distance measure methods, we propose hesitant fuzzy C-means algorithm and Pythagorean fuzzy C-means algorithm, especially for images in real scenarios, we propose hierarchical interferometric fringes identification method based on classic fuzzy clustering, oil well dynamometer cards identification combining fuzzy clustering method based on hesitant fuzzy C-means algorithm, fuzzy clustering method oriented on lithology identification based on Pythagorean fuzzy C-means, and collect real dataset to complete methods validation. On the one hand, this thesis will provide novel idea and method for fuzzy clustering method application in image segmentation, satisfy real industry requirements. On the other hand, it will enrich related theory of unsupervised learning in artificial intelligence.