大数据环境下图像特征提取算法与应用研究

CASIA OpenIR > 毕业生 > 硕士学位论文

	大数据环境下图像特征提取算法与应用研究
	于廷照
	2016-05-27
学位类型	工程硕士
中文摘要	随着移动设备与社交网络的急速发展，互联网中的图像数据呈现指数级爆炸增长。然而，海量图像数据中存在着大量的冗余信息和复杂特征，极大地影响了图像识别、图像复原算法效率及准确率。图像特征提取旨在从纷繁复杂的图像数据中挖掘其潜在信息，因此，研究大数据环境下有效的图像特征提取算法，具有重要的理论意义和应用前景。本文针对大数据环境下，数据无标签、数据量大、结构复杂的问题，提出半监督模式下多标签学习算法，支持的向量筛选降低样本复杂度，特征变换矩阵降低样本维度；针对彩色图像通道间特征提取不一致的问题，提出联合通道内外颜色特征图像复原算法，消除伪彩色，降低模糊效应；针对传统卷积神经元网络复杂度高，高层语义无法解释的问题，提出离散小波卷积网络目标识别算法，降低网络训练时间，提高特征表达能力。以上特征提取算法用于图像识别和图像复原，有效克服了存储空间大、计算缓慢的问题，在标准数据集与主流算法的对比，识别准确率和图像信噪比均获得显著提高。本文主要工作和贡献如下： 1. 提出了一种半监督的多标签学习算法。针对样本无标签的问题，利用样本相似性衡量矩阵，增强学习置信度；针对样本数量大的问题，利用支持向量筛选样本，降低样本数量；针对数据维数高的问题，利用线性变换，降低样本维度。选择九组国际公开数据库，通过与九种主流算法对比实验，实验表明算法在平均准确率上有提升。 2. 提出了一种联合多通道彩色图像复原算法。基于单通道内全局变分特征的颜色相关性，削弱伪彩色效应；基于多通道间梯度特征的颜色差异性，去除模糊效应。选择两组国际公开数据集，一组实际采集数据集，通过与八种主流算法对比实验，实验表明算法在彩色图像信噪比指标上有明显提升。 3. 提出了一种离散小波卷积网络行人再辨识算法。输入层采用离散小波变换，降低样本维度；卷积层采用主成分分析，简化计算复杂度；全连接层采用特征袋模型，降低训练复杂度。选择国际公开通用数据集，通过与十五种算法对比实验，实验表明算法在rank1、rank10、rank20指标上准确率显著提高。
英文摘要	With the rapid growth of mobile devices and social networks, the amount of image data on internet has boomed exponentially. However, redundant information and complicate features are widely existed, which is a challenging task for traditional machine learning algorithms and becomes a major obstacle in image recognition and image restoration. There is an urgent need for exploiting high performance feature extraction algorithm. Feature extraction aims to dig out the essential characteristics from the complex appearance, which has recently become a focus of machine learning research. This research is of great significance for machine learning theory and application. In terms of the large quantity and the complicated structure of big data, we propose a semi-supervised multi-label learning algorithm with joint dimensionality reduction, where support vectors are adopted for sample screening, and a linear transformation matrix is adopted for dimensionality reduction; In terms of the strong rely on assumption and inconsistency of features of the traditional image restoration algorithms, we present a model for color image demosaicking via joint intra and inter channel information, which can eliminate the effect of artifacts and blur. In terms of the fact that current convolutional neural networks consume much training time and the high level semantic can not be interpreted well, we propose a deep convolutional structure involving discrete wavelet transform, which can reduce the time complexity and improve the expression of feature descriptors; These feature extraction algorithms, which are proper in reducing storage memory and speeding up computation, are implemented in image recognition and image restoration. Experiments on standard data sets and comparisons with other state-of-the-art algorithms demonstrate the effectiveness of our algorithm. The contributions of this work are as follows: 1. We propose a new semi-supervised multi-label learning algorithm. In consideration of the data without labels, we utilize a similarity matrix among instances to get a large confidence interval. With regard to the large quantity of instances, we perform support vector machines for data screening. In terms of the high dimensionality of instances, we demonstrate linear transformation for getting a lower subspace. Comparisons with nine state-of-the-art algorithms on nine real data sets show that the algorithm can help to elevate the performance in average precision. 2. We present a model for image demosaicking via joint intra and inter channel information. Features of color correlation is based on intra channel total variance, which can help to reduce artifacts, and features of color difference is based upon inter channel gradient, which can contribute to compensate blur. Experiments compared with other state-of-the-art algorithms on both standard and our own data sets demonstrate the high effectiveness of our model. 3. We propose a discrete wavelet transform involved deep convolutional structure for person re-identification. Discrete wavelet transform is involved in input stage for reducing sample complexity, Principle Component Analysis is involved in convolutional stage for reducing computing complexity, and Bag of Features is involved in fully connected stage for reducing feature complexity. Experiments compared with other state-of-the-art algorithms on standard datasets demonstrate that our algorithm achieves the highest performance.
关键词	特征提取大规模数据稀疏表示半监督学习
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/11834
专题	毕业生_硕士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	于廷照. 大数据环境下图像特征提取算法与应用研究[D]. 北京. 中国科学院研究生院,2016.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
大数据环境下图像特征提取算法与应用研究.（14141KB）	学位论文		限制开放	CC BY-NC-SA