让计算机像人一样能够识别视野范围中的目标，感知所处的环境，理解周围的世界一直以来是计算机视觉技术致力于解决的问题。但是在现实环境中总是充斥着各种干扰信息，例如噪声、遮挡、光照变化、几何失真、视角变化、平移、旋转、仿射变换、尺度变化、场景布局、物体外观等干扰因素给有效地完成计算机视觉任务带来了困难。机器学习方法的发展为计算机视觉中相关关键问题的解决起到了推动作用。利用机器学习的方法可以有效地降低干扰因素影响、减轻模型过拟合问题、解决小样本问题等。借助机器学习中相关算法，技巧，以及框架，本文致力于解决计算机视觉中的相关问题。此外癌症生物信息学作为计算机和生物的交叉学科，获得了越来越多的研究者的关注。利用机器学习方法解决癌症中的相关问题具有重要意义。通过机器学习方法解决例如乳腺癌亚型分类，癌症分期分类等问题，不仅可以帮助研究者了解乳腺癌亚型的不同功能，癌症分期的功能变化等，而且可以辅助医生进行癌症诊断，为患者提供更加准确的治疗方案。 本文的主要研究内容和主要贡献如下， 1.人脸识别中的线性子空间学习的关键问题包括，1)人脸图像的不相似度度量问题。2)人脸图像空间结构保持问题。本文针对这两个问题进行了具体的研究，对于第一个问题，本文提出采用改进的EMD距离来度量人脸图像blocks上的pHOG直方图的不相似性。这样做的目的是一方面可以降低改进的EMD计算的复杂性，另一方面使得改进的EMD对遮挡、遭受、以及其他干扰因素具有较强的鲁棒性。针对于第二个问题，将人脸图像划分为多个blocks，在每个block上提取pHOG直方图。本文提出了基于改进的EMD度量的LPP算法。该算法通过累加两张人脸图像中对应blocks上pHOG直方图不相似性来作为两张人脸图像的不相似性，然后通过计算K近邻来构建近邻图G〓，然后完成人脸图像子空间的学习。为了更有效地利用人脸图像的空间结构信息，本文提出基于改进的EMD度量的BSLPP算法。该算法首先计算人脸图像上对应blocks上的子近邻图G〓，然后合并这些子近邻图获得最后的近邻图G〓，从而完成对人脸图像的子空间学习。 2.视觉词典学习的两个主要问题是，1)视觉词典的大小，也就是视觉单词或者码字的个数如何确定。2)如何保持图像中局部特征的空间结构信息。本文针对于这两个问题进行研究，对于第一个问题，本文提出了一个数据驱动的两层视觉词典学习框架，将视觉词典分成属性层和细节层。在属性层，通过贝叶斯非参数模型自动地确定潜在属性，不仅可以确定潜在属性的个数，而且能够根据数据规模确定混合模型的复杂度，从而减轻过拟合问题。对于第二个问题，本文采用金字塔BOF算法，保留了图像中近似的几何空间对应关系。得益于对这两个问题的研究，本文提出的框架在具有挑战性的15类场景类别数据集上获得了较高的准确率。 3.“打电话”，“骑马”和“跑步”等人类行为的静态属性驱动着本文研究基于静态线索的识别方法。本文构建了一个非顺序卷积神经网络(NCYN)模型用于静态图像行为识别。该模型的优点包括，1)在迁移学习思想的指导下，使用预训练的VGG16来初始化卷积层模块的权重。2)采用数据增强减轻过拟合；使用全局平均池化(GAP)使模型变得更轻，从而提升模型的泛化能力。3)为基线CNN模型和NCNN模型设计了一种端到端的结构，使得提出的CNN模型在只有CPU的PC上进行训练变得可能。4)本文提出的NCNN模型具有非顺序拓扑结构，使得该模型可以分别学习并行分支的空间和通道特征。最后，本文还提出了一种模型组合方法，将本文提出的基线CNN模型和NCNN模型的预测进行权重整合，得到最后的预测。 4.高权重的差异表达基因(DEGs)具有高判别性和生物学重要性。因此基于高权重DEGs，本文分别为乳腺癌亚型构建了相应的二分类器。乳腺癌各亚型的二分类器评估结果验证了高权重DEGs的有效性。本文采用高权重DEGs分别为每个亚型的对照组和实验组构建基因共表达网络，并分析了对照组和实验组相互作用机制的不同。基于这个发现，本文提出一种新的通路富集分析方法，基于基因共表达网络的富集分析(PEGCN)，从而在基因共表达对是否被激活或抑制的层面上进行通路富集分析，PEGCN方法将会对乳腺癌亚型的生物学功能变化进行一定程度的解释。 5.基于基因表达数据的癌症分期分类可以在基因表达值的变化上解释癌症不同分期的功能变化，从而有利于发现和揭示癌症发展和进化的机制。但是目前来讲，用基因表达数据进行癌症分期分类的效果还不理想。本文猜测其中一个原因可能是一维的基因表达值信息缺乏强有力的区分性。因此分析基因之间的交互作用机制将会丰富单个样本的信息。本文提出分别为癌症分期的对照组和实验组构建显著性差异共表达网络(SDCN)，通过SDCN网络来展示对照组和实验组显著差异的交互网络结构。本文提出采用具有稀疏性的显著差异共表达网络(SDCNS)模型来提取稀疏的特征，从而增强癌症分期分类模型的有效性。SDCN和SDCNS的有效性在于将一维的基因表达值信息扩展到了二维的显著性差异共表达基因对上，因此增加了特征的判别性。基于对照组和实验组显著差异的SDCN结构，本文提出了一个使用基因共表达对的富集分析方法(PEUCGP)，通过该方法分别进行广义上调和下调共表达富集分析。通过本文提出的癌症分期框架，完成了癌症不同分期分类模型的构建，并且在一定程度上对癌症的进化机制进行了有效的解释。 关键词：机器学习，计算机视觉，癌症生物信息学，模型过拟合，深度卷积神经网络

 The computer vision technologies alwasys aim to enable computers, like human beings, to recognize the objects in the visual field, perceive the environment, and understand the world around them. However, in the real world, there always exists various kinds of interference information, such as noise, occlusion, illumination changes, geometric distortion, view change, translation, rotation, affine transformation, scale change, scene layout, object appearance and other interference factors, which have brought certain difficulty in effectively conducting the computer vision tasks. The development of machine learning methods promotes the solution of key problems in computer vision. Machine learning methods can effectively reduce the influence of interference factors, alleviate the model overfitting, and solve the model training issues faced with small samples. With the help of relevant algorithms, techniques and frameworks in machine learning, this paper is devoted to solving the corresponding problems of computer vision fields. In addition, cancer bioinformatics, as an interdisciplinary subject of computer and biology, has attracted more and more researchers' attention. It is of great significance to apply machine learning methods to solve the related problems of cancer classification. To solve the problems such as breast cancer subtype classification and cancer staging classification based on machine learning can not only help researchers recognize the different functions of breast cancer subtypes, functional changes in cancer staging, etc., but also assist doctors to conduct cancer diagnosis and provide more accurate treatment for patients The main contents and contributions of this paper are as follows. 1.The key issues realted to linear subspace learning in face recognition include: 1) The issue of dissimilarity measurement of face images. 2) The issue of spatial structure preservation of face images. In this paper, we carry out specific research on these two problems. For the first problem, this paper proposes to adopt the improved EMD metric to measure the dissimilarity of pHOG histograms on blocks of face images. On the one hand, it aims to reduce the computational complexity of the improved EMD, and on the other hand, it strives to show the strong robustness to occlusion, exposure, and other interference factors of the improved EMD. For the second problem, face images are divided into blocks, and pHOG histogram is extracted on each block. In this paper, the improved EMD metric for LPP is proposed. The algorithm first uses the sum dissimilarity betweem pHOG histograms on the corresponding blocks in the two face images as the dissimilarity of the two face images, then constructs the adjacency graph G〓 by calculating the K-nearest neighbors, and finally completes subspace learning for face images. In order to better utilize the spatial structure information of face images, the improved EMD metric for BSLPP is proposed in this paper. The algorithm first calculates the sub-adjacency graph G〓 on the corresponding blocks in the face images, then combines these sub-adjacency graphs to obtain the final adjacency graph G〓, and finally conducts the subspace learning for face images. 2.The two main issues of visual dictionary learning are: 1) How to determine the size of the visual dictionary, that is, how to determine the number of visual words or codewords. 2) How to preserve the spatial structure information of local features. In this paper, we conduct the specific study on these two problems. For the first problem, this paper proposes a data-driven two-layer visual dictionary learning framework, which divides the visual dictionary into the attribute layer and detail layer. In the attribute layer, the latent attributes are automatically determined by the Bayesian nonparametric model, which not only can automatically determine the number of latent attributes but also can determine the complexity of the mixture model according to the data scale, which can alleviate the overfitting problem. For the second problem, this paper utilizes the pyramid BOF algorithm to preserve the approximate geometric spatial correspondence of each image. Benefiting from the study of these two issues, the framework proposed in this paper achieves high accuracy on the challenging fifteen scene categories dataset. 3.The static nature of human actions such as “Phoning”, “RidingHorse” and “Running” drives us to study the static clues-based recognition methods. In this paper, a non-sequential convolutional neural network (NCNN) model is constructed for action recognition in still images. The advantages of the model include: 1) Adopt the pre-trained VGG16 to initialize the weight of the convolutional layer module under the guidance of the transfer learning. 2) Use data augmentation to mitigate overfitting; use Global Average Pooling (GAP) to make the model lighter, thus aims to improve the generalization ability of the model. 3) Design an end-to-end structure for the Baseline CNN model and NCNN model, making it possible to train the proposed CNN models on a PC with single CPU. 4) The NCNN model proposed in this paper has a non-sequential network topology, which enables the model to learn the spatial and channel features of parallel branches separately. Finally, this paper also proposes a model ensemble method, which integrates the predictions of the Baseline CNN model and the NCNN model with weighted coefficients to get the final prediction. 4.High-weight differentially expressed genes (DEGs) have high discriminatory and biological importance. Therefore, based on high-weight DEGs, this paper constructs a corresponding binary classifier for each breast cancer subtype. The evaluation results of the binary classifier for each breast cancer subtype have validated the effectiveness of high-weight DEGs. In this study, we construct gene co-expression networks using high-weight DEGS for the control group and the experimental group of each subtype respectively and analyze the differences in the interaction mechanism between the control group and the experimental group. Based on this discovery, a novel pathway enrichment analysis method, pathway enrichment based on gene co-expression networks (PEGCN), is proposed to analyze the pathway enrichment at the level of whether gene co-expression is activated or inhibited. PEGCN will give a reasonable explanation for the biological function changes of breast cancer subtypes to a certain extent. 5.Cancer staging classification based on gene expression data can explain the functional changes of different cancer stages by examing the change of gene expression values, thus can facilitate to discover and reveal the mechanism of cancer development and evolution. However, at present, the performance of cancer staging classification based on gene expression data is not satisfactory. One of the reasons may be that the information related to one-dimensional gene expression values lacks strong discriminatory power. Therefore, analyzing the interaction mechanism between genes will enrich the information of a single sample. In this paper, a significantly differential co-expression network (SDCN) is constructed for the control group and the experimental group, respectively. SDCN network is used to reveal the significant difference of the interaction network structure between the control group and the experimental group. In this paper, we propose the significantly differential co-expression network with sparsity (SDCNS) to extract sparse features, so as to enhance the effectiveness of cancer staging classification model. The validity of SDCN and SDCNS lies in that it extends the information of one-dimensional gene expression values to two-dimensional significantly differential coexpression gene pairs, thus it promotes the discriminability of features. Based on the SDCN structures with significant differences between the control group and the experimental group, this paper proposes an enrichment analysis method, pathway enrichment using co-expressed gene pairs (PEUCGP), which is used to perform generalized up-regulated and down-regulated co-expression enrichment analysis. Through the cancer staging framework proposed in this paper, the classification models of different cancer stages have been constructed, and the evolutionary mechanism of cancer has been effectively explained to some extent. Keywords: Machine learning, Computer vision, Cancer bioinformatics, Model overfitting, Deep convolutional neural network