Ensemble of Receptive Fields for Training Central-Focused Convolutional Neural Networks
Wenzhao Shao, Po Yang, Yun Yang · 2019
Translation is a data augmentation method widely used in the image classification tasks. We analyze the mechanism of translation and discover that the central area of the images is more likely to be selected as convolutional neural network's input. Inspired by the structure of human retina, we propose the hypothesis that the central area of the image contains more significant information than the marginal one. Comprehensive experiments are presented to prove that hypothesis and reach the conclusion that the receptive field is nonuniform. The central part of the image that is always selected by translation is called the focused area. Motivated by the demand to take use of different focused area and thus take use of different receptive fields, we propose a novel training mechanism that integrate different focused areas in one training process. Our method consists of several stages, each has its own focused area and learning rate, and achieves considerable result in the experiments. We call the integration of focused areas the ensemble of receptive fields.