基于深度置信网络的图像分类方法研究

With the rapid development of visual content analysis and understanding, image classification has attracted growing attentions. The new research on deep architecture, which composed of many hidden layers of neural networks, argues that the deep learning methods which based on deep architecture have the potential to improve the performance. This project studies the image classification problems based on classical deep learning method, deep belief networks. The main content of the study including, 1) iterative deep belief networks are proposed based on previous research results, discriminate deep belief networks, which optimizes the deep architechture and modifys the training methods at the same time. 2) fuzzy deep belief networks are proposed, which inherits the powerful abstraction ability of deep architecture and powerful fuzzy classification ability of fuzzy sets. 3) apply discriminate deep belief networks in handwritten Chinese character recognition mission, which integrates the abstraction ability of deep learning method and discriminative ability of exponential loss function, uses deep architecture for coarse classification and modified quadratic discriminant function for fine classification. We can improve the image classification ability of deep architecture continually, and improve the handwriting Chinese character recognition accuracy by studying. The exploration of this project is important for the application of deep learning methods in image classifcation and handwriting recognition.

随着可视化数据分析和理解的需求越来越大，图像分类的研究日显重要。最新在由多个神经网络的隐藏层组成的深层架构上的研究成果表明，基于深层架构的深度学习方法的性能还有很大的提升空间。本项目基于经典的深度学习方法- - 深度置信网络，研究图像分类问题。研究的主要内容包括，1) 基于前期的研究成果- - 区分深度置信网络方法，优化深层架构，改进训练方法，研究迭代深度置信网络方法；2) 将区分深度置信网络的抽象能力和模糊集的区分能力相结合，研究模糊深度置信网络方法；3) 将区分深度置信网络方法应用到手写中文识别中，将深层架构的抽象能力和指数损失函数的分类能力相结合，使用深层架构进行粗分类，然后使用改进的二次分类函数进行细分类。通过研究，进一步提升深层架构的图像分类能力，提高手写中文识别的正确率。本项目研究对探索深度学习方法在图像分类和手写中文识别中的应用具有重要意义。

为提升图像分类和手写识别的性能，在深度置信网络基础上，实现了一系列分类方法。本项目在迭代深度置信网络、模糊深度置信网络、混合深度置信网络、缺失深度置信网络、提取深度置信网络以及基于深层架构的手写识别方法方面进行了研究。取得的成果包括：实现了迭代深度置信网络，使用无监督学习和监督学习方法迭代训练每一个模块；实现了模糊深度置信网络，继承了深层架构优异的数据抽象能力和模糊集优异的分类能力；实现了混合深度置信网络，混合使用限制玻尔兹曼机和卷积限制玻尔兹曼机构建网络提升分类能力；实现了缺失深度置信网络，来解决存在数据缺失图像的分类问题；实现了提取深度置信网络，将量子计算与模糊技术融入深层架构，来提升图像分类性能。最后，项目组基于上述研究成果实现了基于深层架构的手写识别方法，来解决无约束手写中文识别问题，通过深层架构提升了手写中文的分类性能。本项目的研究对探索深度置信网络与其他方法的融合应用具有重要意义，实现的基于深层架构的手写识别方法，有一定的实用价值。