【作者】 张祥合；

【导师】 任露泉；

【作者基本信息】 吉林大学 ， 仿生科学与工程， 2012， 博士

【摘要】 图像理解是以图像为研究对象，以人类知识为核心，主要研究图像中有什么目标、目标之间的相互关系、图像是什么场景以及如何应用场景的一门科学。其主要研究方向有场景中目标的分类与识别、场景描述与理解两个方面。其中，目标识别是为了对场景进行更好的解释，是场景描述与理解的基础，具有主动性。场景描述与理解为目标识别提供先验信息，可以指导目标识别。从本质上说，图像理解源于计算机视觉，但也融合了人工智能和认知学领域的知识信息，是与计算机视觉、人工智能和认知学相互关联又相互独立的研究领域。目标识别作为图像理解的基础，已应用于身份确定、智能交通管理、机动车辆检测等领域。目标识别的任务是使计算机具有对视觉图像进行复杂背景下检测、识别目标并进行分类的能力。但目前用于目标识别的模板匹配、边缘检测和统计等方法都缺乏生物视觉系统的智能性。自20世纪90年代开始，王守觉院士研究分析了人类认识事物的过程，总结出了同源样本的连续性规律，并称之为同源连续性原理，然后将此原理应用到模式识别中，提出了仿生模式识别方法，并由此形成了高维空间仿生信息学理论，该理论的提出为如何利用计算机解决形象思维问题开辟了一个新途径。本文从研究生物视觉系统的工作机理出发，结合王守觉院士关于高维空间仿生信息学的相关理论，构建相应的计算机模型，研究了图像理解中的目标识别与分类问题，针对不同场景提出了多种目标分类与识别方法。实验结果表明，本文提出的复杂场景中目标的特征提取模型以及在此基础上提出的仿生识别与分类方法有效可行。本文的主要工作总结如下。（1）应用神经生物学知识，对物体识别中涉及的各个脑区分别建模，提出了一种基于视皮层不变性识别机制的类表象仿生构建方法；并利用真实图像组成的训练集和测试集对模型和算法进行测试。结果表明，该方法可以在这些样本上建立有效的深层表征并在其上完成类表象的构建。（2）利用心理学中表象的形成与知觉信息处理过程具有很强相似性的特点，本文将表象系统与视知觉系统建立连接，找到识别活动中各层细胞的活动状态，认为表象的构建来源于识别活动中习得的多个神经元集群的活跃状态，并把其作为表象构建的来源和基础——表象基。通过对识别活动中视皮层各脑区细胞的活动状态进行分别建模，并在其上进行高层次的模式记忆和分析，实现了表象基的提取。（3）结合神经生物学关于视皮层研究的最新成果，将物体识别中的不变性问题引入表象式深层表征的目标范畴确定中。利用腹部视觉通路中细胞的“形状调谐”以及“范畴调谐”的特点，对视皮层细胞的特异性和不变性进行了权衡。（4）利用Bag of words模型构造了Codebook高维空间，利用高维形象几何理论，将图像向量映射成Codebook空间中的样本点，并根据仿生信息学的同源连续性原理，使用BP神经网络分类器找到不同类别目标在Codebook空间中的最佳划分，从而提出了一种复杂场景中特定目标的仿生识别与分类方法；并基于该方法设计了不含隐层和含有一个隐层的BP神经网络仿生分类器，通过实验探讨了影响分类器识别准确率的若干因素及其变化规律，并将本文方法与传统方法得到的实验结果进行对比，证明了论文中提出方法的有效性。（5）在研究了人类视觉系统处理机制的基础上，首先利用方向梯度描述符（HOG）建立了图像的鲁棒表示；然后，根据人类视觉系统的并行处理机制和仿生信息学理论，提出了基于HOG+SVM和HOG+RBPNN的两种人体行为仿生识别与分类方法。利用针对识别与分类方法的评价指标对本文方法进行了评价，最后，与目前常用方法进行了比较，结果表明，在针对静态图像中人体行为的分类与识别效果方面，本文方法对差别较大的行为的识别效果好于常用方法，对相似行为的识别效果还有待于进一步提高。综上所述，本文基于人类视觉系统和高维空间仿生信息学理论，提出的复杂场景中目标的仿生识别与分类算法，可为其他目标识别方法和高维空间仿生信息学的研究提供借鉴。更多还原

【Abstract】 Image Understanding (IU) is to study images based on human knowledge, whichfocuses on what are objects in the image, the relationship between objects, what is scene inthe image and how to apply the scene. It mainly includes two parts: object classification andrecognition in a scene, and scene description and understanding. Object recognition has theinitiative and is the base of scene description and understanding, which is to make scene beexplained well. Scene description and understanding provide the apriori information forobject recognition, which can guide the object recognition. In essence, IU derives fromcomputer vision, while combining the knowledge of artificial intelligence and cognitivescience. It has the close relation with computer vision, artificial intelligence and cognitivescience, but it is independent of them. As the base of IU, object recognition has been appliedin many fields, such as identity determination, intelligent transport management, vehicledetection, etc. Object recognition aims to make computer capable of detecting visual image,recognizing object and classifying object in complex scene. At present, there are manymethods of object recognition, such as template matching, edge detection, statistic methodand so on. However, they lack of intelligence compared with biological visual system.Since the1990s, Professor Shoujue Wang has studied the recognition process of human,and summarized the principle of continuity in homologous samples, called as the principle ofhomology—continuity. He applied this principle to pattern recognition and put forwardbiomimetic pattern recognition and the theory of Multi-Dimensional Space BiomimeticInformatics, which provides a new route for solving imagery thinking problems by computer.In this paper, based on the mechanism of biological visual system and combining the theoryof Multi-Dimensional Space Biomimetic Informatics, object classification and recognition inIU are studied, and the corresponding computer models are constructed. Several methods ofobject recognition are proposed for different scenes. Experimental results show the feature extracting model of object in complex scene and the method of biomimetic classification andrecognition proposed in this paper are effective and feasible.(1) Based on the knowledge of neurobiology, the model of each cerebral regioninvolved in recognizing is established respectively and a biomimetic construction method ofcategory mental imagery based on the recognition mechanism of visual cortex of humanbrain is proposed. The model and algorithm are tested using training set and test setcomposed of real images. The results show that this method can establish valid deeprepresentation of these samples, based on which the biomimetic construction of categorymental imagery can be achieved.(2) The mental imagery in psychology has strong comparability with cognitiveinformation processing. In this paper, mental imagery system is linked with visual cognitivesystem, and the activity status of cells in each brain region in recognition process is found. Itis considered that the establishment of mental imagery derives from the activity status ofmany nerve cells acquired in recognition process, which is the source and base of mentalimagery construction, called as imagery basis. By the activity status of cells and their modememory and analysis in depth, imagery basis is extracted.(3) Combining the hot research results of neurobiology on visual cortex, the invariancein object recognition is introduced to determine the object category of deep presentation ofmental imagery. The shape tuning and category tuning characteristic of cells in ventral visualpathway are employed to weigh the specificity and invariance of visual cortex cells.(4) Codebook multidimensional space is constructed using Bag of words model. Imagevectors are projected to sample points in Codebook multidimensional space. Based on theprinciple of homology continuity in multidimensional space biomimetic informatics, theoptimal classification of different kinds of objects in Codebook multidimensional space isobtained using neural network classifier. A biomimetic classification and recognition methodof specific object in complex scene is proposed. Experiments are performed to investigatethe factors influencing the performance of classifier and their variation. The experimentalresults obtained by this method are compared with those obtained by the traditional methods, and results show the method proposed in this paper is valid and feasible.(5) The robust representation of image is established by Histogram of Oriented Gradient(HOG). According to the processing mechanism of human visual system and the theory ofmultidimensional space biomimetic informatics, two biomimetic classification andrecognition methods of human behavior are proposed, which are based on HOG+SVM andHOG+RBPNN respectively. The methods proposed in this paper are evaluated and comparedwith other commonly used methods. Results show that, for the classification and recognitionof human behavior in still image, the proposed methods have better performance inrecognizing different kinds of behavior, but the performance in recognizing similar behaviorsstill needs improvement.To sum up, a biomimetic classification and recognition method of specific object incomplex scene is proposed based on the processing mechanism of human visual system andthe theory of multidimensional space biomimetic informatics, which can provide a referencefor the research on other object recognition methods and multidimensional space biomimeticinformatics.更多还原