基于蚁群的显著性算法研究

发布时间：2018-05-17 02:29

本文选题：蚁群算法 + 人眼视觉系统　；参考：《北京邮电大学》2016年硕士论文

【摘要】：显著性检测是一种始于生物学方面,于20世纪90年代被引入计算机领域的图像和视频处理方法。根据人类视觉选择注意方式,显著性检测方法可以分为两类:一类是纯数据驱动,独立于任务的自底而上的显著性检测方法,另一类是受意识支配依赖于任务的自顶而下的显著性检测方法。显著性检测对于图像和视频的自动化处理非常重要,它现在已经应用到了图像分割、图像自适应压缩、图像识别、图像非真实感绘制等众多图像处理研究领域,通过图像显著性信息的引导可以更加精准高效地进行图像处理工作。虽然图像显著性检测技术研究已经有了相当不错的成果,但是视频显著性检测的方法还很有限,随着图像和视频处理智能化发展趋势的需求,以及更多领域的使用和普及,显著性检测技术还有很大的发展前景本论文考虑了蚁群优化算法这一基于蚂蚁觅食的生物行为启发的算法和视觉显著性检测之间的关系,以蚁群算法为基础,结合传统的显著性检测思想和算法,提出了一种基于蚁群优化算法的新的显著性检测模型,并对图像和视频的显著性检测方法进行了实验测试,对比其他的算法进行了性能和优缺点评估,同时思考并提出了改进措施或改进方向。本论文中的算法按照检测目标分为两类:非压缩域图像和视频的显著性检测算法研究与压缩域视频的显著性检测算法研究。对于非压缩域图像和视频的显著性检测算法研究,本论文从不同的特征提取方法入手,对基于不同特征及不同尺度的显著性检测结果进行了分析比较,并从中选出效果最优的模型进行实验测试,与其它经典显著性检测算法进行性能对比分析;对于压缩域视频的显著性检测算法研究,本文采用直接从视频压缩码流中提取特征的方法,从残差变换系数中提取亮度、色度以及纹理等空域特征,从运动矢量中直接提取时域特征,对不同的特征进行显著性检测,并采用适应人眼视觉系统的融合方法进行结果融合,获得最终显著性图,同时提出多尺度显著性检测算法,同样与经典显著性检测算法进行了性能对比分析,验证了本论文算法的有效性及可靠性。
[Abstract]:Salience detection is a kind of image and video processing method which started in biology and was introduced into computer field in 1990s. According to the human visual choice of attention, salience detection methods can be divided into two categories: one is purely data-driven, a bottom-up, mission-independent salience detection method. The other is the top-down salience detection method, which is controlled by consciousness. Salience detection is very important for the automatic processing of image and video. It has been applied to many research fields of image processing, such as image segmentation, image adaptive compression, image recognition, image non-realistic rendering and so on. Image processing can be carried out more accurately and efficiently through the guidance of image saliency information. Although the research of image salience detection technology has made quite good achievements, the methods of video salience detection are still very limited. With the development trend of intelligent image and video processing, and the use and popularization of more fields, There is still a great prospect for significance detection. This paper considers the relationship between ant colony optimization, a biological behavior heuristic algorithm based on ant foraging, and visual salience detection, which is based on ant colony algorithm. In this paper, a new salience detection model based on ant colony optimization algorithm is proposed, and the salience detection method of image and video is tested experimentally by combining the traditional salience detection idea and algorithm. Compared with other algorithms, the performance, advantages and disadvantages are evaluated, and the improvement measures or directions are proposed. The algorithms in this paper are divided into two categories according to the detection target: the significance detection algorithm of image and video in uncompressed domain and the salience detection algorithm in compressed video domain. In this paper, we analyze and compare the salience detection results based on different features and scales by using different feature extraction methods to study the salience detection algorithms in uncompressed domain images and video. And select the best model for experimental test, and compare the performance with other classical salience detection algorithms. For compressed video salience detection algorithm research, This paper adopts the method of extracting features directly from video compression bitstream, extracts spatial features such as brightness, chromaticity and texture from residual transform coefficients, extracts temporal features from motion vectors, and detects the salience of different features. The fusion method adapted to human visual system is used to fuse the results and obtain the final significance graph. At the same time, a multi-scale salience detection algorithm is proposed, and the performance of the algorithm is compared with that of the classical salience detection algorithm. The validity and reliability of the algorithm are verified.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】