基于视觉注意机制的目标检测方法与应用研究

发布时间：2018-03-14 23:19

本文选题：目标检测　切入点：视觉注意机制　出处：《西南科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：目前,移动机器人的目标检测作为一项多学科综合的复杂技术,现已渗透到军事、工业、生活等各个领域,得到国内外研究者的广泛关注。移动机器人的目标检测以图像处理、分析和理解为基础,随着图像成像技术的迅速提升,在获得超高分辨率图像的同时,也带来了图像信息冗余的问题,这为图像后期处理(如目标检测等)增加难度。而反观人类视觉系统则具有异常突出的信息筛选能力,能迅速选择进入人眼的重要信息并做出反应,这种具有选择性和主动性的生理和心理活动被称为视觉注意机制。本文对视觉注意机制的相关原理及方法进行了深入研究,并将研究结果运用在目标检测中,实现场景图像目标的快速自动检测。本文主要的研究内容和成果如下:(1)本文通过研究视觉注意生物学理论并分析现有的各显著性检测算法优缺点,结合特殊环境目标检测的具体应用,提出对IT算法进行改进,使之适用于具体目标检测。改进算法从底层特征显著性分析,选取对目标感兴趣区域提取最有利的若干特征进行计算,使得目标感兴趣区域的提取更加准确完整,弥补了原IT算法在求取目标时轮廓信息缺失的不足。由于在特殊环境中采集的图像含有噪音,本文提出了一种将多重中值滤波与曲波变换相结合的混合去噪算法进行降噪预处理。实验表明该方法较好地降噪并保护边缘,改善图像的视觉质量。在进行降噪预处理的基础上,本文利用改进IT算法进行感兴趣区域提取,并进行目标检测,而目标检测的正确检测率达89%,证明基于视觉注意机制的特殊环境目标检测方案可行。(2)本文利用人工智能中的随机森林算法训练好的显著性回归器映射每个区域特征向量的显著值而获取显著图,使得目标检测过程更加智能化。在求取区域特征时,本文针对具体的应用场景(商场物品检测),选取了对目标感兴趣区域计算有利的区域对比度、背景度以及属性等特征。实验证明本文算法虽然在用时上稍比其他几种对比算法用时长,但显著性检测结果明显远远优于其他几种常见或经典的算法。同时,目标检测的正确检测率达92%,证明可以利用视觉注意机制进行商场物品检测。
[Abstract]:At present, the target detection of mobile robot, as a complex and multidisciplinary technology, has been infiltrated into military, industrial, life and other fields, and has been widely concerned by researchers at home and abroad. The target detection of mobile robot is based on image processing. On the basis of analysis and understanding, with the rapid improvement of image imaging technology, the problem of redundant image information is also brought about by obtaining ultra-high resolution images. This increases the difficulty of image post-processing (such as target detection, etc.), while the human visual system, on the other hand, has an exceptionally prominent information screening ability to quickly select and respond to important information that enters the human eye. This selective and active physiological and psychological activity is called visual attention mechanism. In this paper, the related principles and methods of visual attention mechanism are deeply studied, and the results are applied to target detection. The main contents and results of this paper are as follows: 1) this paper studies the biological theory of visual attention and analyzes the advantages and disadvantages of the existing salient detection algorithms. Combined with the specific application of target detection in special environment, this paper proposes to improve the IT algorithm and make it suitable for specific target detection. The extraction of the region of interest of the target is more accurate and complete by selecting the most favorable features for the region of interest extraction. It makes up for the lack of contour information in the original IT algorithm. Because the images collected in the special environment contain noise, In this paper, a hybrid de-noising algorithm combining multiple median filtering and Qu Bo transform is proposed. Experiments show that the proposed method can reduce noise and protect edges. Improve the visual quality of the image. Based on the pre-processing of noise reduction, this paper uses the improved IT algorithm to extract the region of interest and detect the target. The correct detection rate of target detection is 89, which proves that the special environment target detection scheme based on visual attention mechanism is feasible. 2) in this paper, we use the random forest algorithm in artificial intelligence to train a significant regression to map each region special. The salience value of the eigenvector to obtain the salience graph, The process of target detection is more intelligentized. In order to obtain the regional features, this paper selects the region contrast that is favorable to the area of interest to the target, aiming at the specific application scene (shopping mall item detection). The experimental results show that the proposed algorithm is much better than other common or classical algorithms, although it is slightly longer in time than other comparison algorithms. The correct detection rate of target detection is 92, which proves that the visual attention mechanism can be used to detect shopping mall items.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【相似文献】