少量训练样本下基于深度学习的图像分类研究

Image classification with a small number of training samples is a difficult problem in the fields of computer vision and artificial intelligence. At present, most of the related works make up for the shortage of data by data augmentation and domain adaptation, and use regularization technologies to alleviate the overfitting of the model. However, they can not still meet the requirements which image classification with a small number of training samples has for model ability. This project is based on the deep convolution neural networks (DCNN). It will study the key theoretical problems in image classification with a small number of training samples, including the learning of features that are discriminative between classes and are common within classes, the construction of network structure which could alleviate the overfitting of the model, and the ensemble classification method that can ensure the accuracy and diversity of base learners. The main contents include: First, learning the relationship between images based on DCNN and studying the learning of features that are discriminative between classes and are common within classes; Second, constructing a network structure that can reduce the function space of DCNN and alleviate over-fitting; Third, constructing a base classifier based on the proposed feature learning method and network structure, and studying the image ensemble classification method of DCNN under a small number of training samples. The study of this project is of great significance to promote the widespread application of machine recognition technology, to break through the bottleneck of small sample learning theory, and to realize artificial general intelligence in China.

少量训练样本的图像分类是当前计算机视觉和人工智能领域的一个难题。目前，相关工作大都通过数据增强和领域自适应的方法来弥补数据量不足的缺陷，并使用正则化技术来缓解模型的过拟合，但仍无法达到少量训练样本的图像分类对模型建模能力的要求。本项目拟基于深度卷积神经网络（DCNN），对少量训练样本的图像分类中（1）类间有区分性且类内有共性的特征学习、（2）缓和过拟合的网络结构的构建、（3）确保基学习器准确而不同的集成分类方法，这三个核心理论问题展开研究。主要包括：基于DCNN学习图像间的关系，研究类间有区分性且类内有共性的特征学习方法；构建缩小DCNN函数空间并能缓解模型过拟合的网络结构；结合提出的特征学习方法和网络结构搭建基分类器，研究少量训练样本下基于DCNN的图像集成分类方法。本项目的开展对推动机器识别技术的广泛应用，为我国率先突破小样本学习理论的瓶颈和实现通用人工智能具有非常重要的意义。

相对于大规模图像分类，基于深度学习的小样本图像分类工作主要关注如何基于少量训练样本提高分类性能，已取得了丰硕成果，但仍然存在一定的改进空间。本项目研究少量图像样本下类间有区分性而类内有共性的特征学习，提出了基于滑动特征向量的小样本图像分类方法，最大限度地利用深层特征信息的同时扩充了类特征空间，从而提高了分类的精度；研究一种缩小DCNN函数空间并缓解模型过拟合的网络结构，提出了面向小样本图像分类的正交分类层，从而学习到一个较大的决策边界和更有判别力的特征表示；研究少量训练样本下基于DCNN的图像集成分类方法，提出了面向小样本图像分类的混合注意力网络和集成分类学习方法InterBoost，从而获得了更多的、具有高辨识度的局部特征，并使得分类器集成后获得了更好的泛化能力。该项目的开展为小样本学习提供了一个重要思路，同时也对图像分类、小样本学习以及神经网络的理论研究具有重要意义。