Improving Efficiency of CNN using Octave Convolution

Abstract

In recent years, the Convolutional neural networks (CNN) has been active in various Artificial intelligence applications as well as computer vision tasks. We suggested an effective technique in this study to decrease the number of duplicates in feature maps of CNN. Proposed a novel convolution scheme Octave convolution (Octconv) to minimize the duplicates in the feature maps and boost the CNNs performance. The principle concept of this method is to separate the Convolutional filters into a higher frequency and lower frequency sections. In this report, we made an attempt for minimizing the spatial redundancy directly from output feature maps of CNN using the following 3 steps: First, divide the channels into higher and lower frequency parts depending on the information of the image using Multi-scale representation. Second, reduce the number of FLOPs from the low frequencies. Third, before sending the output to combine both the higher frequency and lower frequency information of the image. The key purpose of this abstract is to improve CNNs efficiency by reducing spatial redundancy in the feature maps of the convolution layer.