基于类编码的深度网络学习研究

Feature representation is an important topic in pattern analysis. Deep Learning, as an end-to-end feature learning method, has achieved great success in computer vision and pattern classification. Auto-encoder network, as an unsupervised network, has achieved good results in character recognition and image classification. This project is based on the auto-encoder network, and incorporates the label information, which is important and very related to classification, to propose a network of class encoder, including the data or feature based class encoder and its deep structure. Applying to image recognition field, we further propose class encoder regularized deep learning method, multi-modal class encoder network and multi-task class encoder method, and propose fast structure selection method for deep class encoder network with big data. By studying the characteristic of class encoder, we try to improve the discriminative ability and the generalization performance of deep neural network, enrich the deep learning algorithms and improve the accuracy of pattern classification.

特征表达是模式分析中的重要研究内容。深度学习，作为一种端到端的特征学习方式，在计算机视觉、模式分类中取得了成功的应用。自编码网络，作为一种非监督网络结构，在字符识别、图像分类等领域取得了较好的结果。本课题以自编码网络为基础，结合对分类性能至关重要的类别信息，提出并研究类编码神经网络，包括基于数据重建和特征重建的类编码深度网络结构，结合图像识别应用，提出基于类编码正则化的深度网络学习方法、多模态类编码网络和多任务类编码深度网络学习方法，并进一步结合大数据，提出类编码深度网络结构快速选择方法。通过提出一整套类编码网络体系，改进现有深度网络的判别性和推广性，丰富深度学习算法，提高模式分类的准确性。

本课题面向模式分析中的特征提取展开研究，重点研究深度学习中的类编码网络以及深度学习在图像分类等任务中的应用。本课题提出了类编码网络，深度类编码网络以及类编码网络正则化的优化方法。结合图像大数据，在图像预处理，人脸识别，目标再识别，人脸属性分析，图像分类等领域开展了深度学习方法的研究，提出了新的算法和优化目标，提升了相关任务的性能，取得了国际先进的识别性能。本项目成果丰富了深度学习的理论，为深度学习在图像分析领域的应用提供了新思路。