Disentangled convolution for optimizing receptive field

Abstract

Convolutional neural network (CNN) is one of the primary techniques for high-performance image recognition. The convolution operation with small-sized filters is a key ingredient of CNN and the receptive field of the whole CNN is enlarged by stacking lots of convolution layers. The convolution layer, however, is problematic in terms of the receptive field. It provides fixed small receptive field due to the fixed filter size in convolution, which requires us to manually control it in advance. Besides, the larger-sized convolution filters significantly increase computation cost. Thus, in this study, we propose a method to adaptively tune the receptive field of the convolution operation in an end-to-end manner as well as to enlarge the receptive field in a low computation cost. Based on the biological studies and scale-space theory, we can disentangle convolution operation into Gaussian envelope filtering for smoothing and derivative-related filtering, both of which are heterogeneously parameterized. Those two types of filters are jointly optimized in the end-to-end CNN training and the receptive field of the convolution is adequately optimized via learning the Gaussian envelope with a low extra computation cost. The experimental results on image classification tasks demonstrate that the proposed method effectively enlarges receptive field to improve performance.