【摘要】 卷积神经网络通过学习图像的特征表示进行图像分类,然而图像特征中往往存在大量的无关特征甚至是干扰特征,这些特征的学习阻碍了网络性能的提升,因此,如何去除无效特征,强化关键特征成为了CNN图像分类的研究方向之一。注意力机制模仿人脑接受外部信息时只处理重要信息而滤除次要信息的机制有效的解决了这个问题。将注意力机制与CNN结合可以更好的关注图像上下文信息,使网络具有甄别特征中关键特征的能力,提高网络性能。本文从基础的CNN注意力模型出发,分析介绍了不同的注意力模型及其发展方向,从多个角度综合概述了不同注意力模型的优缺点和差异性。最后,本文对这些网络模型做出了分析总结,探讨了CNN注意力模型在图像分类领域存在的一些问题和未来可能的研究方向。更多还原

【Abstract】 Convolutional neural Network classifies images by learning the feature representation of the image, but there are often a large number of irrelevant features and even interfering features in the image features, and the learning of these features hinders the improvement of CNN performance. Therefore, how to remove invalid features and enhance the key features has became one of the research directions of CNN image classification. The Attentional mechanism(AM) effectively solves this problem by imitating the mechanism that only deals with important information and filters out secondary information when human brain receives external information. The combination of attentional mechanism and CNN can better pay attention to the image context information,so that the network has the ability to discriminate key features and improve the network performance. Starting from the basic CNN attention model, this paper analyzes and introduces different attentional models and their development direction, and comprehensively summarizes the advantages and disadvantages and differences of different attention models from multiple perspectives. Finally, this paper makes an analysis and summary of these AM models, and discusses some problems and possible future research directions of CNN attention model in the field of image classification.更多还原