A Down-sampling Method Based on The Discrete Wavelet Transform for CNN Classification

Convolutional neural networks usually have an upper bound for the input size due to a limited hardware resource, which causes the inconsistent size problem that the images are generally too large to match the input requirement. Therefore, an efficient down-sampling method is in need urgently. Existing methods remove a great amount of information when reducing the size, resulting in accuracy degradation. We start with the Grid Channels Stack method and propose a novel down-sampling method which based on the discrete wavelet transform. Our method not only retains more information than the commonly used bilinear interpolation but also suppresses the distortion impact compared to the Grid Channels Stack method. We evaluate the proposed method with the popular ResNet-18 network on the CIFARIOO and the MINI-ImageNet datasets. The improved performance demonstrates the effectiveness of our method, especially for the various compression rate occasions, and we believe it will be a promising down-sampling method for the CNN domain.