Improve light weight convolutional neural networks for dace recognition with multi-style recalibration module

Abstract

With the rapid development of deep learning, many areas have achieved impressive progresses. However, deep learning often suffers from complex computation paradigms and huge memory occupancies, lightweight networks attract gradually increasing attention recently. Various lightweight network models have emerged in recent years, and are also widely used in small mobile devices. Then, there is still a lack of expressive power for lightweight networks compared to deep networks. This study leverages multiple statistics as representative of a feature map and proposes a Multi-Style Recalibration Module of a channel attention mechanism to improve the accuracy of lightweight network model recognition without increasing their storage space. This mechanism is embedded on MobileFaceNet as a study case. We test the new model on LFW dataset for demonstration and find that its occupancy only increases 0.2 MB, while the accuracy is increased by 0.18%.