图像特征提取方法及其应用研究

发布时间：2018-06-18 00:42

本文选题：时频复合加权 + 抗噪柱状图　；参考：《西北大学》2016年博士论文

【摘要】：计算机视觉和图像处理领域普遍存在数据维度高,图像数据类型日益复杂的情形,经典的计算和分析方法对这类图像数据进行分析处理时往往计算代价过高,甚至会完全失效。通常在对高维复杂图像数据进行分析和处理之前,需要对样本数据集进行特征提取操作,提取与样本相关性强的特征点,并且要去除噪声特征点和与样本数据集不相关的冗余特征点,以便为后续的处理工作提供高价值的数据。通常图像特征提取方法所获取的特征点,在表征图像时存在优劣之分,同时特征点的优劣直接影响着实际应用中图像处理的结果。图像特征提取是进行图像配准、图像目标识别、图像检索等应用的关键步骤,从图像数据中提取表达能力强、抗噪能力好的图像特征仍然是图像处理领域的研究难点和热点之一。本文在对常用的特征提取方法进行深入分析的基础上,针对若干特定应用领域的图像特征提取方法进行了深入研究,提出了一系列改进算法,以提高特征提取的有效性和实用性。具体内容如下：1提出了一种约束优化进化的夜间图像时频复合特征提取方法。对多帧夜间模糊图像的时域与频域同时加权处理,是实现深度提取夜间模糊图像特征的先进方法。传统的夜间图像特征提取方法大多只单独提取时域特征,没有对多帧夜间低质量图像的频谱特性进行分析与特征提取。本文方法针对多帧原始图像,分别在频域和时域提取多帧图像之间的相关信息；通过加权处理形成新的图像特征；运用约束优化进化算法对图像特征提取的结果不断进行循环优化,最终达到更好的特征提取效果。2提出了一种基于柱状图特征描述的含噪图像特征提取方法。在基于内容的图像检索方法中,通常将图像的内容表示成柱状图,根据图像柱状图之间的相似性进行检索。由于数码图像中包含噪声,往往使得柱状图变得平滑,图像之间变得更为相似,导致返回结果中图像数量增加,检索准确率降低。为了进一步提高图像检索方法性能,本文提出了一种对噪声不敏感的柱状图特征描述符,并用该特征描述符进行图像检索。将图像中的噪声描述为平稳附加高斯白噪声,并给出相应的柱状图表示。通过随机变量的原点矩定义了柱状图的特征描述符,并且分析了如何应用特征描述符恢复原始图像的柱状图。3提出了一种基于LBP和人眼视觉感觉模型的航空图像特征提取方法,研究非单一目标的航空图像准确检索的问题。由于航空图像特定区域中的目标呈随机性分布,干扰目标与识别目标交错分布。传统的图像检索算法的思路都是以目标为线索,根据特定的目标像素特征,或者关联特征完成图像的检索。干扰排除也以特征对比为主,方法较为机械,在目标众多环境下,排除过程极其复杂,检索效率与准确性都很低。为了避免上述缺陷,本文方法利用局部二进制(LBP)方法进行航空图像的特征提取,并将上述不同种类的特征作为人眼视觉感觉航空图像检索的基础数据。通过建立人眼视觉感觉模型,进行航空图像检索,提高了检索效率和准确性。4提出了一种基于SIFT描述符的图像角点特征提取方法。该方法使用最近距离比的方法来获得初始匹配特征描述符对,有效减少异常值的影响。此外,利用尺度方向的联合限制寻找假匹配SIFT描述符对。并利用随机采样一致性原则删除异常值,提高图像角点特征配准精度。5为实现对图像的实时边缘检测,研究了基于嵌入式多DSP的实时边缘检测系统。设计了基于3个DSP两两相连的图像边缘检测运算模块,并通过FPGA对图像进行预处理来提高检测效率,保证了边缘检测的实时性。采用改进的万有引力边缘检测算法,降低噪声对检测结果的影响,提高对细节信息的检测准确度。
[Abstract]:In the field of computer vision and image processing, the data dimension is high and the image data type is increasingly complex. The classical calculation and analysis methods often have high cost and even complete failure when analyzing and processing this kind of image data. Usually, before the analysis and processing of the high dimensional complex image data, the sample needs to be matched. This data set extracts the feature extraction operation, extracts the feature points with strong correlation with the sample, and removes the noise feature points and the redundant feature points which are not related to the sample data set so as to provide the high value data for the subsequent processing work. At the same time, the advantages and disadvantages of feature points directly affect the results of image processing in actual applications. Image feature extraction is the key step in image registration, image target recognition, image retrieval and so on. The extraction of image features from image data is strong, and the image features with good noise resistance are still the difficult and hot topics in the field of image processing. Firstly, based on the in-depth analysis of the commonly used feature extraction methods, this paper makes a thorough study of the image feature extraction methods in certain application fields, and proposes a series of improved algorithms to improve the effectiveness and practicability of feature extraction. The specific contents are as follows: 1 a night map with constrained optimization evolution is proposed. The time frequency complex feature extraction method is an advanced method for the time domain and frequency domain weighting of multi frame nocturnal blurred images. It is an advanced method to extract the features of the nocturnal blurred image. Most of the traditional night image feature extraction methods only extract the time domain features alone, and do not analyze the spectrum characteristics of the multi frame low quality images. Feature extraction. This method extracts the related information between multi frame images in frequency domain and time domain for multi frame original images. A new image feature is formed by weighted processing, and the result of image feature extraction is continuously optimized by constrained optimization evolutionary algorithm, and a better feature extraction effect.2 is proposed at the end. In the content based image retrieval method, the content of the image is usually expressed as a columnar graph and retrieved according to the similarity between the image histogram. Because of the noise in the digital image, the histogram becomes smooth and the image becomes more similar. In order to further improve the performance of the image retrieval method, this paper presents a feature descriptor of the histogram that is insensitive to noise, and uses the feature descriptor to retrieve the image. The noise in the image is described as a stationary additional Gauss white noise, and the corresponding column is given. The feature descriptors of the histogram are defined by the origin moment of random variables, and the histogram of the original image which is used to restore the original image by the feature descriptor.3 is analyzed. An aerial image feature extraction method based on the LBP and human visual sense model is proposed, and the problem of accurate retrieval of the non single target aerial images is studied. Due to the random distribution of the target in the specific area of the aerial image, the interference target and the recognition target are interlaced. The traditional image retrieval algorithm is based on the target as the clue, and completes the image retrieval according to the specific target pixel features, or the associated features. The interference elimination is also dominated by the feature contrast, and the method is more mechanical and in the eye. In a large number of environment, the removal process is extremely complex and the retrieval efficiency and accuracy are very low. In order to avoid the above defects, this method uses the local binary (LBP) method to extract the characteristics of the aerial images, and takes the different kinds of features as the basic data of the visual sense of the human eye sense. Through the establishment of human visual sense, the visual sense of human eye is established. An image retrieval method is used to improve the retrieval efficiency and accuracy..4 proposes an image corner feature extraction method based on the SIFT descriptor. This method uses the nearest distance ratio method to obtain the initial matching feature descriptor pair, effectively reducing the effect of the abnormal value. In addition, the joint restriction of the scale direction is used to find the false. Matching the SIFT descriptor pair, and using the random sampling consistency principle to delete abnormal values and improve the registration accuracy of the image corner feature.5 in order to realize real-time edge detection of images, the real-time edge detection system based on embedded multi DSP is studied. A graph image edge detection operation module based on 3 DSP 22 is designed, and a FPGA pair diagram is used. In order to improve the detection efficiency and ensure the real-time performance of the edge detection, an improved universal gravitational edge detection algorithm is adopted to reduce the impact of noise on the detection results and improve the accuracy of the detection of details.
【学位授予单位】：西北大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】