Neural Network Compression via Low Frequency Preference

Network pruning has been widely used in model compression techniques, and offers a promising prospect for deploying models on devices with limited resources. Nevertheless, existing methods merely consider the importance of feature maps filters spatial domain. In this paper, we re-consider characteristics propose novel filter method that corresponds to human visual system, termed Low Frequency Preference (LFP), frequency It is essentially an indicator determines based relative low-frequency components across channels, which can be intuitively understood as measurement “low-frequency components”. When map more than other maps, it considered crucial should preserved during process. We conduct proposed LFP three different scales datasets through several achieve superior performances. The experimental results obtained CIFAR ImageNet dataset demonstrate our significantly reduces size FLOPs. UC Merced show approach also significant remote sensing image classification.