深度学习理论及在图像识别中的应用研究

项目名称： 深度学习理论及在图像识别中的应用研究

项目编号： No.61273364

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 黄雅平

作者单位： 北京交通大学

项目金额： 81万元

中文摘要： 深度学习（Deep Learning）理论是机器学习领域的研究热点，其核心思想在于建立一种具有多层次结构、非线性特性的学习模型。与传统的浅模型（Shallow Model）相比，它能够学习复杂变化的特征，推广能力强，有望解决Transfer Learning等人工智能领域的根本性问题。但是，深度学习理论及算法复杂度较高，并且在图像识别等视觉任务中，与浅模型相比，性能并不突出。因此，本项目从图像识别任务出发，拟深入探索深度学习理论，主要研究：(1)、在生物视觉信息处理机制的启发下，构建符合人类感知系统的深度学习模型，并研究快速收敛算法；(2)、提出具有超完备能力的快速学习算法，解决小样本学习的难题；(3)、将深度学习模型应用到大规模的图像识别任务中，研究半监督学习算法从而提高性能；(4)、针对基于视频的图像识别任务，引入时间相干性准则，研究视频数据的深度学习算法，争取取得实际应用成果。

中文关键词： 深度学习；特征提取；图像识别；机器视觉；神经网络

英文摘要： Deep learning has gained significant interest as a way of building hierarchical representations in machine learning. Deep architectures are composed of multiple levels of non-linear operations, and can learn complex and high-vary function and obtain good generalization compared with shallow architectures. It is more promising that deep architectures may be beneficial to solve some fundamental problems in AI, such as transfer learning. However, current deep learning approaches always suffer from computational complexity due to non-convex optimization. Moreover, they do not work well in vision tasks, e.g., image classification, in contrast to some state-of-art shallow architectures. Therefore, this project aims to present a novel deep learning model and then apply the model to image classification tasks. The main works include four aspects: (1). developing a novel bio-inspired deep learning architecture and proposing the corresponding fast-convergence algorithm; (2). learning overcomplete representations to deal with the situation where the number of outputs of the layer is greater than the number of inputs; (3). combining unsupervised and supervised strategies to boost the performance for large-scale image data set; (4). employing temporal coherence principle to learn more robust features from a video. The perfor

英文关键词： Deep Learning；Feature Extraction；Image Classification；Machine Vision；Neural Networks