Yusuke UCHIDA
Takayoshi YAMASHITA
This paper surveys the convolutional neural network (CNN) for image classification tasks, which is a rapidly developing area in the neural network field. A history of six representative CNNs since AlexNet that appeared in 2012 is reviewed while important components used in those models are explained. Based on the ResNet that was identified to be a completed model up to 2017 at the time of the survey, many models proposed so far are uniquely categorized from the following viewpoints; improvement of the Residual module that characterizes ResNet, improvement of the macro ResNet architecture like module stacking, improvement of generalization performance, and faster inference.
A significant point of this paper is the comprehensive verification of accuracy and processing speed for 10 typical models and 3 types of data sets. Evaluating different models fairly is surprisingly difficult and cannot be achieved without proper parameter selection and model learning by trusted researchers. Furthermore, the deep insights can be derived only by authors with wide knowledge.
Choosing a CNN model is essential for research and development making this of interest for many readers. As described above, since this paper has a very high validity and contribution in this field as a survey of CNN models, it is highly evaluated as a paper worthy of the Society's Best Paper Award.
