FD-SSD: An improved SSD object detection algorithm based on feature fusion and dilated convolution

作者：

Highlights：

• A new multi-scale information enhancement network for small object detection is proposed.

• The semantic information of the shallow feature map is improved by the muliti-layer feature fusion module. Through the multi-branch residual dilated convolution module, the original resolution of the feature map is kept and the context information of the feature map is improved. Besides, deformable convolution is used to fit the shape of small objects.

• Experiments on the PASCAL VOC2007 dataset and MS COCO dataset prove the effect of the proposed network.

摘要

•A new multi-scale information enhancement network for small object detection is proposed.•The semantic information of the shallow feature map is improved by the muliti-layer feature fusion module. Through the multi-branch residual dilated convolution module, the original resolution of the feature map is kept and the context information of the feature map is improved. Besides, deformable convolution is used to fit the shape of small objects.•Experiments on the PASCAL VOC2007 dataset and MS COCO dataset prove the effect of the proposed network.

论文关键词：Small object detection,Multi-layer feature fusion,Multi branch residual dilated convolution,Context information enhancement

论文评审过程：Received 20 March 2021, Revised 6 July 2021, Accepted 2 August 2021, Available online 17 August 2021, Version of Record 23 August 2021.

论文官网地址：https://doi.org/10.1016/j.image.2021.116402