【作者】 张林；

【导师】 李平； 张宇；

【作者基本信息】 浙江大学 ， 导航、制导与控制， 2015， 硕士

【摘要】 在视觉研究领域,目标识别是一个很热门的研究课题,同时也是很有挑战的任务。然而人类大脑视觉皮层的独特信息处理机制,使其具备快速准确高效的目标识别能力。在过去的数十年中,很多视觉算法的提出,改善了目标识别的性能,但是现在的任何一种算法都未能超越人类大脑。因此,构建能够模拟人类大脑视觉皮层的目标识别机制,或者尽可能与人类视觉相匹敌的人工系统,一直都是具有吸引力但也很困难的目标。设计与构建鲁棒的目标识别算法,将会有益于诸多领域与应用,比如安全监控,机器人导航,医学图像理解等。本文主要围绕目标识别与图像分类算法展开研究。本文既借鉴神经生物学与认知生理学目前的研究成果,又考虑人类大脑的生物学习机制,结合HMAX (Hierarchical Model and X)模型与极限学习机,设计与构建新颖的生物记发式图像分类算法网络。本文的主要研究内容包括：第一,传统视觉与生物视觉中的目标识别问题的研究,以及目标识别与图像分类问题的数学描述与解决方案。目标识别分为两种情况,鉴别与分类。从生理学和计算方面同时考虑目标识别问题的解决方案与策略,并且指出解决目标识别问题的两个重要阶段：构造包含足够多的有效信息的目标图像特征,与选择恰当的分类器对目标图像特征进行分类。这为目标识别与图像分类算法的设计提供总体理论支撑。第二,生物启发式图像分类算法网络的设计与实现。本文设计的生物启发式图像分类算法网络包括五个层级,分别是S1层、C1层、S2层、C2层与H层,即S1-C1-S2-C2-H。其前四个层级主要注重设计与构建特征结构描述,基于灵长类动物的视觉皮层生理学数据构造图像特征数据向量。网络的H层主要关注人类大脑的生物学习机制,通过极限学习机分类器来实现,对图像特征数据向量进行分类。同时,本文尝试将生物启发式特征构建机制与学习机制融合来构建新算法,为深入研究目标识别引入新思路与新视角。第三,生物启发式图像分类算法网络的实验结果与分析。本文选取三个图像数据集,设计了四个组别的图像分类实验,分别是识别准确率验证,参数敏感性分析,混淆矩阵计算,以及隐藏结点比较。通过实验验证与分析,本文提出的图像分类算法网络,具有较好的识别准确率性能以及快速的学习速度,同时也具有潜在的工程应用前景。总而言之,本文提出的生物启发式图像分类算法网络,在模仿人类大脑的识别与学习能力的算法研究上向前迈进了一小步。本文的研究工作在尝试搭建计算机视觉与神经科学之间沟通的桥梁。更多还原

【Abstract】 Object recognition has been a popular area of intense research, and is also a very challenging task in computer vision, while human vision with unique processing me-chanism has the ability to recognize objects rapidly, accurately, and effortlessly. Al-though the performance of the object recognition has been improved with the vision algorithms, which have been proposed during the last decades, none of these algo-rithms available today can surpass the performance of the human brain. Thus, build-ing an artificial system that emulates object recognition in cortex or matches with human vision as closely as possible has always been an attractive but elusive goal. To design and build robust object recognition would be beneficial for many fields and applications, such as security surveillance, robot navigation, clinical image under-standing, etc.This paper mainly focuses on object recognition and image categorization algo-rithms. In accordance with the current achievements of neurobiology and neurophysi-ology, and considering the biological learning mechanism, we combine the Hierar-chical Model and X (HMAX) model and the extreme learning machine (ELM), and then construct a novel biologically inspired feed forward network for the image cate-gorization. The main work and contributions of this paper are summarized as follows:Firstly, the research of object recognition in computer vision and biological vision, and the problem and strategy of object recognition and image categorization. Object recognition includes two cases, identification and categorization, respectively. Its strategy and solution has been discussed in the view of physiology and computation. There are two main steps in object recognition based on image. One is to generate an effective feature representation of an image that includes sufficient information, and the other is to classify the feature representation with a proper classifier. This provides general theoretical supports for the design of object recognition and image categoriza-tion algorithms.Secondly, the design and implementation of the biologically inspired image ca-tegorization algorithm network. The proposed image categorization algorithm network consists of five layers, namely, S1 layer, C1 layer, S2 layer, C2 layer and H layer, that is, S1-C1-S2-C2-H. The previous four layers focus on the design of feature represen-tation structure, and build high-level image feature representations based on physio-logical data about the mammalian visual pathways. The H layer at last pays attention on biological learning mechanism of the human brain, which is implemented with ELM, and acts the role of classifying the image feature representations. Meanwhile, the proposed network tries to combine the biological feature building mechanism and the biological learning mechanism together, which provides a new train of thought for in-depth study on object recognition.Thirdly, the experimental results and analysis for the biologically inspired image categorization algorithm network. In this paper, four groups of experiments on the three images datasets have been performed, which are predicting accuracy verifica-tion, parameters sensitivity analysis, confusion tables calculation, and hidden nodes comparison, respectively. Experimental results show that the proposed network has good performance with fast learning speed. And it shows a potential application prospect in the engineering.Conclusively, the proposed biologically inspired image categorization algorithm network, seems a little step towards human brain alike recognizing and learning, and is also an attempt to bridge the gap between computer vision and neuroscience.更多还原