Feature refinement for image-based driver action recognition via multi-scale attention convolutional neural network

作者：

Highlights：

• We create a real driver action dataset R-DA.

• We employ multi-scale convolutional block with maximum selection unit to adaptively learn multi-scale information from different receptive fields of convolutional kernels.

• We incorporate an attention mechanism to learn spatial saliency and channel saliency for feature refinement.

• We validate the effectiveness of our proposed MSA-CNN framework on multiple driver action recognition datasets and report the comparisons with existing solutions.

摘要

•We create a real driver action dataset R-DA.•We employ multi-scale convolutional block with maximum selection unit to adaptively learn multi-scale information from different receptive fields of convolutional kernels.•We incorporate an attention mechanism to learn spatial saliency and channel saliency for feature refinement.•We validate the effectiveness of our proposed MSA-CNN framework on multiple driver action recognition datasets and report the comparisons with existing solutions.

论文关键词：Driver action,Attention mechanism,Maximum selection unit,Fine-grained

论文评审过程：Received 30 January 2019, Revised 2 October 2019, Accepted 8 November 2019, Available online 15 November 2019, Version of Record 19 November 2019.

论文官网地址：https://doi.org/10.1016/j.image.2019.115697