自组织学习及其在汉字识别和图象分析中的应用

【作者】 邓达；

【导师】 余英林； 陈国评；

【作者基本信息】 华南理工大学 ， 通信与电子系统， 1995， 博士

【摘要】 随着生物学、心理物理学对人类感知系统的深入了解，神经网络研究中以非监督学习、自组织的方式从现实世界提取特征并映射归类的机制引起了越来越大的关注。较之于已得到广泛应用的监督式学习算法，自组织学习具有好的生物可行性，能够在缺乏指导信息的情况下形成对输入数据的有效表征。 本文基于人的早期视觉模型，采用自组织学习的若干算法，做了以下三个方面的工作： 首先，我们尝试将一种线性主分量分析网络用于图象分割的实验，并参考视觉模型，引入了非线性机制大幅度地提高了算法收敛速度。 其次，我们采用一组Gabor滤波器作为视觉接受场，应用于手写体汉字识别中的特征提取过程。对得到的特征进一步运用自组织映射，并辅以监督式的学习过程构造特征映射码本，最后得到了较好的识别效果。 针对纹理图象分割的问题，我们也围绕Gabor滤波器及自组织映射，提出了一个综合纹理的频率信息和方向选择性的编码方案和计算模型。 最后，我们反过来研究以自组织学习机制获得类似Gabor滤波器的视觉感受场的问题。为此，我们提出了一个新的“自组织Hebbian学习”算法，并构造了一个多分辨率的图象金字塔用于学习。我们的实验表明，学习得到的感受场模板在空间频率和方向上都有一定的选择性。更多还原

【Abstract】 Human efficiency in real world signal processing and pattern recognition tasks has been enthralling and far-away from modern computer practice for ages. While biological and psychophysical endeavors are exploring the complex mechanism in human perception, a number of artificial neural network models have been built to enhance the performance of computer systems in real world understanding and adaptation.Among these network models implementation with supervised learning algorithms, such as Back-propagation, has been quite successful so far, yet substantial draw-backs limit their wider application in solving complex problems. In this thesis we pay special interest in a school of self-organized constructions. These models, bearing more biological plausibility, are oriented to feature extraction, clustering, and topological mapping of input data, and all these characteristics are achieved without guidance of a teacher. Another issue of interest is the early vision model, of which, thanks to biological and psychophysical discoveries, have become relatively well-known. The main achievement in this thesis, are based on both of the two topics we mentioned above.We first test the ability of a PCA network, based on unsupervised Hebbian learning, in image segmentation. Nonlinear mechanism is introduced to enhance the convergence speed of network training.Apart from conventional methods such as structure analysis, we try to tackle the problem of handwritten Chinese chiracter recognition from an early-vision point of view. A set of Gabor filters, analogous to simple cells in visual cortex. are used to extract local orientational and spatial frequency features occurred in the character image. These features are further piped into a Self-Organized Mapping (SOM) process to be clustered, and the mapping result is learned by a supervised algorithm LVQ so that good recognition result is obtained.We also proposed a network model for texture segmentation, also based on SOM and Gabor filtering.Finally, a computational mode! for the problem of visual receptive field learning is constructed. A multi-resolution image pyramid, originally used for image analysis, is applied to provide spatial frequency information into the learning process, which we further strengthen with a new Self-Organized Hebbian Learning algorithm. As a result, receptive fields with both orientation and spatial frequency selectivity, similar to Gabor filters, are learned and tested.更多还原