【作者】 王宇新；

【导师】 郭禾；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2012， 博士

【摘要】 图像识别是模式识别领域中的一种典型应用,也是计算机应用领域中一门崭新的技术,它利用计算机对图像进行处理、分析和理解,以识别各种不同模式的目标和对象。场景语义分类和场景中的目标识别是这个领域中的热门研究课题,其中场景中的目标识别针对目标类别问题,让计算机识别出场景中有什么物体,在哪儿出现；而图像的场景分类是场景描述和理解的重要研究内容,目标是根据给定的一组语义类别对图像数据库进行标注,如海岸、山脉、森林、街道等,它为引导目标识别等其他场景内容的进一步理解提供了有效的上下文信息,具有重要的指导作用。本文重点关注图像场景中的目标识别和场景分类问题,在对图像进行色彩、纹理、形状和局部不变性特征等特征提取的基础上,引入了高层语义特征,主要是利用图像中的对象与对象间、对象与场景间的空间关系来进行图像识别,对图像内容的理解上升到了对象及其空间关系的理解,并重点关注图像分块机制、图像中的视觉特征分布对分类效果的影响。具体来说首先以传统词袋模型为基础,根据同类场景图像具有空间相似性的特点,提出一种用于图像场景分类的空间视觉词袋模型,在场景分类和目标识别两类任务上都取得了明显的识别率提升。同样基于同类场景图像具有的空间相似性,以及图像块间内容的相关性,提出一种基于PageRank过滤噪音图像块,再结合先验知识判断测试图像块的场景信息,最后采用图像块语义投票机制的场景分类方法,取得了很好的识别效果。其次,在分析图像特征分布的基础上,提出一种视觉词汇重要性评估方法,在降低图像特征直方图维数的同时,提高了分类准确率。同时通过分析Census变换直方图和PACT特征的分布特点,在CT特征中引入隐含阶梯边缘模板实现了特征降维,在识别率相当的情况下改进了算法执行的效率。再次,将图像识别技术成功应用于一个基于图像识别技术的海冰图像监测系统中,用于测量海冰厚度、密集度和流速等参数,并实现了多种解决实际问题的图像识别方法。同时,提出了一种结合CUDA并行编程模型和OpenMP并行技术的图像特征提取与匹配加速策略。最后,通过深入研究图像模板匹配策略,基于模糊理论,提出了一种用于源码中的设计模式识别的模板匹配方法,有效提高了匹配的准确性。更多还原

【Abstract】 Image recognition is a typical field of pattern recognition and a new technology in the area of computer application. It uses computers for image processing, analysis and understanding, to identify a variety of different patterns of target objects. In this area, scene semantic classification and object recognition in the scene are popular research topics. Object recognition is for the object category problem and let the computers identify which object is in the scene and where it is; the scene classification is a very important research filed of scene description and understanding, and its goal is to give a set of semantic classes annotation, such as coast, mountains, forests, streets, etc., for the images in database, which plays an important role in guiding object identification and can provide effective context information for further understanding of other content in the scene image.This paper focuses on the problem of object recognition and scene classification, based on the extraction of images’color, texture, shape and local invariant feature, high-level semantic features are introduced, mainly the use of the spatial relationship between objects and relationship between object and the scene in the image, for image recognition. The understanding of the content of the image has been increased to the understanding of the image objects and their spatial relationship, and especially the influence of image block mechanism and the distribution of visual features in the image to classification effect is focus on.Specifically, at first a model named bag of spatial visual words is proposed, which is based on traditional bag model and spatial similarity of the image in same scene classification, and made significant recognition rate improvement in both scene classification and object recognition tasks. Similarly, based on the spatial similarity between images and content correlation between image blocks, A new scene classification method is presented, in which the noise blocks are filtered by PageRank algorithm firstly, and then prior knowledge is combined to determine the test image blocks’scene information, finally voting mechanism is used for semantics scene classification. The final recognition results achieved is good.Secondly, an importance assessment method of visual vocabulary is presented based on the analysis of the distribution of image features, which can reduce the dimension of image features histogram while improving classification accuracy. At the same time through the analysis of spatial distribution characteristics of Census Transform histograms and PACT, the implicit step edge template is introduced to spacial PACT feature, and the implementation efficiency of the algorithm is improve greatly with a equivalent recognition rate.Then, a sea ice monitoring system for measuring sea ice thickness, density and velocity and other parameters based on image recognition technology is designed and implemented, and a variety of image recognition solutions for solving practical problems are proposed. At the same time, we propose an acceleration strategy combined with CUDA parallel programming model and OpenMP parallel technology to accelerate the image feature extracting and matching process.Finally, through in-depth study of the image template matching strategy, we propose a template matching method based on fuzzy theory for recognizing design patterns from the source code and improve the matching accuracy effectively.更多还原