基于层次深度网络混合模型的图像识别技术研究

Large-scale image recognition is a challenging hot research topic, and it has received increasing attention from both academia and industry. However, due to the complexity of image contents, it is very hard for the traditional classification methods to accomplish such an ambitious task by using the low-level visual features. In the recent years, the deep neural network can significantly improve the accuracy of image recognition. But, the following two questions are not well resolved: 1) Only single deep network is used for learning of features and classifiers, and the correlation among different categories are not fully utilized to learn the features and classifiers; 2) The hierarchical classifier cannot well tackle the overlapping between foreground and background as well as the error propagation problem during the feature learning process. Therefore, in this project, we propose to study the mixture of hierarchical deep networks for large-scale image recognition in an end-to-end manner, which focuses on solving (a) the learning problem of the hierarchical multi-task model on top of deep learning and the multi-level tree structure, (b) the prediction problem of the hierarchical visual attention model for deep multi-task network, and (c) the mixture learning issue of multiple networks. This proposal would improve the semantic modeling and the image recognition performance, and help to build more practical image recognition system, and further promote the development process of the visual cognition mechanism in the artificial intelligence era.

大规模图像识别是一个极具挑战性的热点研究方向，受到学术界和工业界越来越多的关注。然而，由于图像内容的复杂性，传统基于底层视觉特征的分类方法已经难以完成这项艰巨的任务。近年来，虽然深度学习技术能够显著地提高大规模图像识别的准确率，但是仍存在如下两方面问题：1）仅采用单一网络参与特征学习与分类器训练，不能充分地利用不同类别间的相关性；2）层次分类器不能很好地解决特征学习中前背景重叠和误差传递问题。因此，本项目拟以端到端的深度学习模型为基础，研究基于层次网络混合模型的图像识别算法，着重解决在深度学习和层次树结构上层次网络模型学习问题、深度网络层次视觉关注模型预测问题以及多网络模型混合学习问题。本项目的研究可以显著地提升图像语义建模水平和识别的准确性，有助于构建实用的图像识别系统，能够推进人工智能时代视觉认知机制的发展进程。

图像识别是人工智能、计算机视觉以及多媒体分析等领域的研究热点。本项目在深度学习框架下，结合相关领域知识，以层次结构建模为基础展开图像识别相关算法研究：1）在层次特征表达和分类器学习方面，研究了基于树结构的层次深度网络模型，利用类间相关性关系，提高了算法的效果并降低了时空复杂度，实现了从粗到细的可解释图像识别；2）在前景估计方面，探索了层次化视觉识别机制，提出了层次视觉关注模型以学习粗糙类和细粒度类的视觉差异，实现了个性化的层次特征表达，提升了结果的可解释性和视觉识别效果；3）在混合模型方面，针对单一算法模型表达力弱和扩展性差的问题，研究了深度网络模型的混合特征学习方法，通过任务空间分组、知识蒸馏以及互补学习等策略，进一步提高了层次化模型的视觉识别能力。在项目执行过程中，项目组在特征表达、视觉关注、混合模型设计等相关理论、算法和技术方面积累了丰富的经验，可以为相关研究及应用提供理论和技术支持。在本项目支持下，项目组培养了多位人工智能、计算机视觉以及多媒体分析等领域的人才，部分相关研究成果已申请专利或发表在国内外高水平国际期刊及会议上，并逐步开展了与工业界的产学研合作。