一种基于原型学习的多示例卷积神经网络

A Prototype Learning Based Multi-Instance Convolutional Neural Network

卷积神经网络是一种全监督的深度学习模型,其要求样本类标完整.在样本类标缺失等弱监督的实际应用中,卷积神经网络的应用受到了极大的制约.为解决弱标记环境下的多示例学习问题,该文提出了一种新的多示例深度卷积网络模型.该模型引入了一种新的原型学习层.该层使用基于原型度量的算法,实现了示例特征至包特征的映射,从而使网络能够在包的层面给予类标信息,进而完成整个模型的学习过程.该文首先在肺癌病理图像细胞分类的问题中,验证了该网络的性能.实验表明,相较于传统基于手工图像特征的方法,该文所提出的方法在准确率方面约有12%的提升.相较于卷积神经网络结合传统多示例学习的方法,所提出的方法在各项指标上同样取得了更好的效果.此外,在自然图像分类数据集GRAZ-02上,所提出的方法相较于目前最优的算法也取得了相当的效果.

Convolutional neural network is a fully supervised deep learning model. It requires that the labels of samples are fully provided. In weakly supervised applications where labels of samples are partly provided, the usage of convolutional neural networks is greatly limited. To solve the weakly supervised multi-instance learning problem, a new multiple instance convolu- tional neural network is proposed. The proposed model introduces a new prototype learning layer into the network. The prototype learning layer uses a prototype based metric method to transform instance features into bag features. The network therefore can use label information of bag and learning the whole model in a compact process. The network is firstly tested on a lung cancer cell pathology image classification dataset. Results show, compared with hand designed image feature based methods, the proposed method achieved an improvement of about 12% in accuracy. Compared with convolutional neural network and multi-instance learning combined methods, the proposed method also achieved better results on all the evaluation criterion. Besides, the method is also tested on a natural image classification dataset （GRAZ-02）. Comparable result is achieved by the proposed method compared with the state-of-the-art method.