【作者】 任桢；

【导师】 王科俊；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 随着数字信息技术的发展及广泛应用,数字图像和数字视频的数量增长迅速。图像分类任务正是在这样的一个前提下提出并发展起来。图像分类研究任务主要由预处理,特征提取和分类三个主要环节构成,每个环节对图像的分类效果都有重要的影响。本文从这一着眼点出发,对图像分类各环节的关键技术进行逐一分析,针对各个环节的处理任务,提出以下方法:(1)针对光照对图像分类的影响,提出了自动截断拉伸的快速多尺度Retinex方法(TWMSR)。通过自动截断拉伸处理修正了多尺度Retinex方法(MSR)从log空间映射回灰度空间后,拉伸受少数极值点影响而造成的失真;提出一种窗口无关快速均值滤波算法,用于替代MSR方法中的高斯环境函数,提高了MSR方法的运算速度。TWMSR方法与多种亮度归一化和彩色常化方法进行去光照对比实验,证实该方法在亮度归一和彩色常化上具有最佳性能。(2)为了去除噪声对图像分类的影响,提出梯度均变双边滤波图像去噪(GSBF)方法。通过分析图像的构成特性提出了图像均质判定规则,构建了梯度均变去噪方法;通过分析梯度均变方法和双边滤波方法性能,提出GSBF方法,实现了去噪和细节保留的平衡。将GSBF方法与多种去噪方法从主、客观的角度和特征稳定性角度进行对比实验,证实了GSBF方法能更有效地去除噪声,提高特征的稳定性。(3)在特征区域获取环节,提出最大分布熵多尺度小波显著特征区域获取方法。通过采用最大分布熵确定待选取的小波特征点数量,有效地控制了特征的分布性;通过引入多尺度Log空间,实现了不同尺度小波特征区域的获取。将文中提出的小波区域获取方法与其它特征区域求取方法进行对比实验,从尺度、模糊、旋转、光照、视角五种图像特性变化的角度,根据重复性标准进行了比较,结合对这些算法所求得特征区域的相关性评价,提出联合特征区域求取方法。通过实验证明,该联合特征区域求取方法满足了特征区域提取的多样性。(4)通过对不同图像特征描述子的描述特性分析,构建了一种联合特征描述子,该联合特征描述子由具有信息互补性的4种描述子构成。针对联合特征描述子的特性,提出一种基于打分制的Recall-precision特征描述子评价方法,将该联合描述子与其它描述子从图像特性改变的角度进行了实验比较,证实该联合描述子能够更稳定的描述图像特征区域。通过对联合描述子进行性能实验分析,确定了联合描述子的融合系数。(5)分析视觉字集在图像分类中的应用,提出一种改进的K均值聚类方法,并用其生成视觉字集。结合概率拉丁语义模型(pLSA)和基于高斯混合贝叶斯两种分类训练模型,对文中所讨论的算法进行综合实验。实验表明,这些环节的改进有效的改善了图像分类效果,也进一步证实了各环节算法在相应图像处理功能上的有效性。更多还原

【Abstract】 Along with the development and abroad application of digital information aquiring techniques, the number of digital images and digital videos has grown enormously. Image classification task is developed under this kind of background and composed of image preprocessing, feature extraction and classifying processing three steps. Each step has some important affects on the final classification results. This paper just focuses on this side, analyzes the key techniques of those steps in image classification in turn. And effect methods have been proposed with each step’s processing tasks:(1)Aim at the light impact for image classification, we propose an automatic truncation stretch and fast multi-scale Retinex method (TWMSR). Through automatic truncation stretch method, repair the distortion of MSR caused by max or min pixels during the process of mapping from log space to gray space; propose a fast window size irrelevant mean filtering method to substitute the Gaussian environment function and improve the computing speed of MSR. Comparing the TWMSR method with several other illumination normalizing and color constancy methods and prove that TWMSR method has better performance in illumination normalizing and color constancy.(2) For the popuse of getting rid of the noise impact for image classification, a gradient symmetrically changing bilateral filtering de-noising method (GSBF) is proposed. By analyzing the characteristics of image composing we propose the decision rules of symmetrically changing of image; combining the gradient symmetrically changing with bilateral filtering method. And then rebuild the filtered image iteratively using rules of symmetrically changing. Compare the GSBF method with some other de-noising methods, and give an analyzing on the stability of features. It proves that GSBF method can remove the noise more effectively and can get more stable features than others.(3)Aim at feature region capturing, a max distributing entropy wavelet-based salient multi-scale region detector (WSMR) has been proposed. It controls the features distributing characteristics efficiently by using max distributing entropy; by introducing multi-scale log space it can get wavelet feature regions with different scales. Evaluate the proposed WSMR and some other feature region detection methods on scale, blur, rotation, light and view angle. Use repeatability criterion to compare those methods and combine them with relative criterion between those features detected by those methods, bring forward the unite feature region detection method. By experiment it has shown that the unite feature region detection method fulfills the diversity of feature regions.(4)Through the analyzing on the expressed information carried by different descriptors, united descriptor method is constructed with several discriptors which have supplement informations for each other. Aiming at the characteristic of united descriptor, a Recall-precision evaluating method based on score rules is proposed. Compare the unite descriptor with some other descriptors and prove that the united descriptor can give the feature region a more stable expression. Further more, according to the comparing of chosen descriptors on several angles of image characteristics changing, the weighted coefficients of unite descriptors are decided.(5)By alalyzing the principle of visual words in image classification, an improved K-means clustering algorithm is proposed to generate visual words. Then give the algorithms analyzing of probabilistic latent semantic analysis (pLSA) model and Bayesian decision classification model based on Gaussian mixture model. By Experiment on those algorithms discussed in the paper and concluded that the improvement in each step can bring up the image classification effect and can also make better performance in corresponding image processing function.更多还原