SRM : A Style-based Recalibration Module for Convolutional Neural Networks

• DOI: 10.48550/arxiv.1903.10829
• Publication Date: 2019-01-01
• Authors: HyunJae Lee; Hyoeun Kim; Hyeonseob Nam
• Keywords: Pooling; Leverage (statistics); Feature (linguistics)

Following the advance of style transfer with Convolutional Neural Networks (CNNs), role styles in CNNs has drawn growing attention from a broader perspective. In this paper, we aim to fully leverage potential improve performance general vision tasks. We propose Style-based Recalibration Module (SRM), simple yet effective architectural unit, which adaptively recalibrates intermediate feature maps by exploiting their styles. SRM first extracts information each channel pooling, then estimates per-channel recalibration weight via channel-independent integration. By incorporating relative importance individual into maps, effectively enhances representational ability CNN. The proposed module is directly fed existing CNN architectures negligible overhead. conduct comprehensive experiments on image recognition as well tasks related styles, verify benefit over recent approaches such Squeeze-and-Excitation (SE). To explain inherent difference between and SE, provide an in-depth comparison properties.