Improved Convolutional Neural Networks by Integrating High-frequency Information for Image Classification

Abstract

Deep convolutional neural networks are powerful and popular tools as deep learning emerges in recent years for image classification in computer vision. However, it is difficult to learn convolutional filters from the examples. The innate frequency property of the data has not been well considered. To address this problem, we find high-frequency information import within deep networks and therefore propose our high-pass attention method (HPA) to help the learning process. HPA explicitly generates high-frequency information via a stage-wise high-pass filter to alleviate the burden of learning such information. Strengthened by channel attention on the concatenated features, our method demonstrates consistent improvements upon ResNet-18/ResNet-50 by 1.36%/1.60% and 1.47%/1.39% on the ImageNet-1K dataset and the Food-101 dataset, respectively, as well as the effectiveness over a variety of modules.