【作者】 吴军；

【导师】 王士同；

【作者基本信息】 江南大学 ， 计算机应用技术， 2012， 硕士

【摘要】 数字图像处理技术随着计算机科学技术的发展和计算机网络的日益普及而迅速发展。图像分类作为其重要分支,为图像的进一步实验和研究提供了很多重要信息。图像分类技术是模式识别在数字图像处理领域的具体运用,其主要目的在于研制出能够代替人类来完成图像分类和识别任务的计算机智能系统。近些年来,人们在这方面做了很多努力,并取得了显著成效。目前主要的图像分类方法有支持向量机(SVM)、基于知识的图像分类法、模糊集理论、人工神经网络方法等。支持向量机在图像的应用也十分广泛,但是它的参数的选择将很大程度上影响到它的性能,而目前还没有十分完备的方法来确定参数;基于知识的图像分类方法主要是以专家的知识和经验为基础的图像分类方法,这种方法引起了很多科研人员的广泛关,但知识和经验大多属于特定的地域和时域,该方法需要解决的是如何使其具有自适应学习能力。模糊集方法有较强的灵活性以及适应性,能很好的处理模糊性问题,但其隶属的函数的选择主要由经验确定,这就导致该方法存在一定的主观性和盲目性。人工神经网络方法利用计算机模拟人类自主学习的过程,具有很强的非线性逼近能力,在信息处理和模式识别的过程中得到了广泛应用。当前,大部分基于模糊系统的图像分类算法都只利用了模糊系统中的正规则,然而负规则往往也提供了很多有用信息。本文着重讨论了正负模糊规则系统在图像分类上的应用。文中将正负模糊规则进行有效结合形成了正负模糊规则系统,提出三种抗噪性能和分类性能较好的正负模糊系统。本文的工作重点是:1、改进了一种正负模糊系统并将其用于图像分类,文中阐述了正负模糊规则相对于正模糊规则以及负模糊规则的优势,详细描述了正负模糊系统的结构、隶属度函数类型以及参数的调整方法,并通过一系列实验证明该方法不仅可以对遥感图像进行分类,而且可以对自然图像进行分类,同时具有较好的抗噪性。2、在此基础上,针对上述正负模糊系统中采用梯度下降法调整参数,导致学习速度慢、容易得到局部最优值等问题,引入了极限学习机(ELM)理论。这不仅有效避免了上述问题,而且在隐含层激励函数无限可微的情况下,可以对输入权值和隐含层阈值任意赋值。文中提出了基于ELM的正负模糊系统的图像分类,并结合遥感图像和自然图像进行了一系列实验。图像分类的结果表明,该方法不仅有较好的分类效果和抗噪效果,而且学习速度快、泛化性能好。3、极限学习机(ELM)是一种单隐含层前馈神经网络(SLFN)。为减少隐含层神经元个数,增加神经网络的灵活性,我们提出了隐含特征空间岭回归(HFSR)并将其拓展到多隐含层前馈神经网络(MLFN)。MLFN的HFSR的隐含层最后一层可以通过求Moore-Penrose伪逆的方法的获得,其他隐含层参数可以任意赋值,同时它具有ELM的学习速度快、算法简洁、网络只需训练一次的优点。文中提出了基于HFSR的正负模糊规则系统的极速图像分类。实验表明,文中方法除了有较好的图像分类效果,也有很好的抗噪性。更多还原

【Abstract】 Digital image processing technology develops rapidly with the development of computer science and technology and the increasing popularity of computer networks. As an important branch of image processing, the image classification provides a lot of important information for further experiments and researches. The technology of image classification is a specific application of pattern recognition in the field of digital image processing; its main purpose is to develop a computer intelligent system which could replace human beings to complete the task of image classification and recognition. In recent years, people made many efforts in this regard and have achieved remarkable results. At present, the main method of image classification are support vector machine (SVM), knowledge-based image classification, fuzzy set theoryand artificial neural network method, etc. Support vector machine is widely used in image classification, but its performance is greatly affected by the parameters, and parameter selection method are not yet comprehensive or complete; Knowledge-based image classification methods based on expert knowledge and experience attract wide attention from many scientific researchers, while most of the knowledge and experience belong to a particular geographical and time domain, and self-learning is the problem need to be resolved; Fuzzy set method which has good flexibility and adaptability can well solve the ambiguity problem, but the choice of membership function mainly relies on experiences with great subjectivity and blindness; Artificial neural network method using computer to simulate self-learning process of human beings and of strong non-linear approximation capability, has been widely used in information processing and pattern recognition.Most of the current image classification algorithms only use the positive rules in fuzzy system, while negative rules also provide a lot of useful information. This paper focuses on the positive and negative fuzzy rule system and its application to image classification. Based on positive and negative fuzzy rule system which effective combinates of positive and negative fuzzy rules, we propose three positive and negative fuzzy system with anti-noise performance and better classification performance. This work will focus on:1, A kind of positive and negative fuzzy system for image classification is improved, which has obviously advantages compared to positive fuzzy rule systerm or negative fuzzy rule systerm. We detailly describe the structure, the type of membership functions and parameters’adjustment methods of positive and negative fuzzy system. A series of experiments show that this method with better noise immunity can classify not only remote sensing image, but also natural images.2, On this basis, since their method heavily suffers from the very slow learning speed for the training and easily falling in local minima of the cost function of the network which is realized using the feedforward neural network model which requires adjusting the weights in the gradient descent way, Extreme Learning Machine (ELM) theory is introduced. It is the effective way that can avoid the above problems and the input weights and hidden layer bias can be randomly assignment when the hidden layer activation function infinitely differentiable. In this paper, the positive and negative fuzzy system based ELM is proposed for image classification and a series of experiments are done on remote sensing images and natural images. Image classification results show that the method has better classification effect and anti-noise effect, and also has fast learning speed, generalization performance.3, The Extreme learning machine (ELM) is a single hidden layer feedforward neural network (SLFN). In order to reduce the number of nodes in the hidden layer and increase the flexibility of neural networks, we propose a hidden feature space ridge regression (HFSR) and expanse to a multi-hidden layer feedforward neural network (MLFN). The parameters in the last layer of hidden layers of the MLFN’s HFSR can be obtaned by Moore-Penrose generalized inverse method, and other parameters in hidden layer can be arbitrarily assigned while it is a simple and fast algorithm just like ELM which only needs to train one time. A positive and negative fuzzy rule system using ridge regression for extremely fast Image classification is proposed in the paper. A series of experiments show that the proposed method has better image classification results, and also has very good noise immunity.更多还原