基于智能优化算法的图像检索技术研究

发布时间：2018-03-04 18:49

本文选题：图像检索　切入点：群体优化算法　出处：《江南大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着科学技术的进步和互联网时代的发展,以及大数据时代的到来,尤其是多媒体技术和数字图像处理技术的广泛应用,导致图像的数据量出现井喷式的增长。与传统的文字、数字等文本的信息表达方式不同,图像蕴含的信息更加丰富和复杂多变,因而针对图片的检索和数据挖掘也更加困难。目前,如何从海量的图像数据库中精准的检索出期望的图像已经成为近十几年来计算机科学领域的研究热点。实现精确地图像检索的关键是图像信息标记方式的选择,近年来,利用图像的颜色、纹理、形状等内容特征来标记图像信息的图像检索技术,即基于内容的图像检索技术(Content Based Image Retrieval,CBIR),已经成为目前图像检索领域的主流发展方向。基于内容的图像检索具有广泛的应用前景和深远的研究价值和商业价值,因而该研究领域引起了相关研究机构和研究人员的高度关注。目前,基于内容的图像检索技术的研究虽然取得了不俗的成果,部分成果甚至已经得到广泛应用,但是还有很多方面存在不足,需要进一步的改善和优化。传统的基于内容的图像检索技术采用单一的图像视觉特征和相似性度量算法进行图像检索,因而无论检索的准度和准度都普遍偏低。针对该问题,本文提出并实现了采用颜色和纹理两种视觉特征以及12种相似性度量算法的基于内容的图像检索方法,并采用QPSO粒子群优化算法进行检索。同时,通过与PSO、CLPSO、SLPSO三种粒子群优化算法的检索效果进行对比选出最优算法,并对最优算法利用GPU加速技术,从而提高图像检索的性能。本文使用的关键技术和理论方法主要包括以下四个方面:(1)由于图像的颜色特征和纹理特征是表达图像内容的最直接的两种视觉特征,因此本文综合使用这两种特征实现基于内容的图像检索。颜色特征方面,基于人类视觉特征采用RGB、HSV、Lab与Gray四种颜色空间,提取图像的颜色直方图与颜色矩特征,并将这些特征进行量化;纹理特征方面,采用该特征描述的两种主要方法,灰度共生矩阵与Gabor图像处理方法以提取图像的纹理特征,并将纹理特征进行量化。(2)利用12种当前常用的相似度距离算法对目标图像和待检索图像库中每一幅图像提取的颜色特征和纹理特征进行度量。(3)通过使用PSO、QPSO、CLPSO、SLPSO四种群体优化算法获得优化特征、相似性度量函数以及权重之间的近似最佳组合,从而使检索效果更加准确和高效。(4)对四种群体优化算法中最优的QPSO算法使用C++AMP技术实现系统的GPU加速,并通过测试对加速的效果进行验证。
[Abstract]:Along with the progress and development of the Internet era of science and technology, and the arrival of the era of big data, especially the wide application of multimedia technology and digital image processing technology, image data lead to blowout growth. With the traditional text information expression of digital text of the different image contains more abundant information and complex and so for the image retrieval and data mining is more difficult. At present, how from the huge image database retrieval in accurate expectations image has become in recent years the computer science research led domain. To realize accurate image retrieval is the key to image information marking way, in recent years, the use of image the color, texture, image retrieval technology to mark the shape of image information content characteristic, namely the content-based image retrieval technology (Content Based Image Retri Eval, CBIR), has become the mainstream of the development direction of the field of image retrieval at present. It has wide application prospect and great research value and commercial value of content based image retrieval, and the research field and attracted the attention of the relevant research institutions and researchers. At present, research on the image retrieval technique has achieved good results based on some results and even has been widely used, but there are still many deficiencies and need further improvement and optimization. Based on the technology used in single image feature and similarity measure algorithm for image retrieval content-based image retrieval and traditional, both accuracy and the accuracy of the retrieval are generally low for. This problem, this paper proposes and implements the two kinds of color and texture feature of visual and 12 similar image retrieval based on content measurement algorithm Using QPSO method, and the particle swarm optimization algorithm for retrieval. At the same time, with PSO, CLPSO, SLPSO three kinds of particle swarm optimization algorithm for retrieval results were compared to select the optimal algorithm, and accelerate technology on the optimal use of GPU algorithm, so as to improve the performance of image retrieval. In this paper, the key technology and theory method mainly includes the following four aspects: (1) the color features and texture features of the image are two kinds of visual features of the most direct expression of the image content, so the use of the two kinds of features for content-based image retrieval. Color characteristics of human visual features using RGB, based on HSV, Lab and Gray four kinds of color space the extraction of image, color histogram and color moment features, and quantified feature; texture feature, the two main methods of the description of the features of the gray level co-occurrence matrix and Gabor image processing. By the method of extracting image texture features, and quantify the texture features. (2) using the 12 kinds of similarity distance algorithm commonly used to retrieve the target image and each image to extract the color and texture features of the image library to measure. (3) by using PSO, QPSO, CLPSO, SLPSO four the group optimization algorithm to obtain optimal feature, similarity measure between the function and the weights of the approximate optimal combination, so as to make the retrieval results more accurate and efficient. (4) the use of C++AMP technology to realize the system GPU acceleration of four group optimization algorithm the optimal QPSO algorithm, and through the test of the effect of acceleration is verified.

【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】