Hao Zhang,Danpu Zhang,Xiaolong Liu,Jianfeng Yang,Pengfei Shi
标识
DOI:10.1117/12.3011350
摘要
We observe representative information of large, medium objects and the vast background overwhelm that of small objects, which leads to poor performance of small object detection. To this end, a new module has been proposed in this paper, named Background and Foreground Attention Maps(BFAM) module, composing of three sub-modules: segmentation, background and foreground attention sub-modules. The segmentation maps positions which have the top strong semantic information in the deep layer of the backbone to the bottom layer maintaining more small object details and mask them to obtain background map and foreground map. Apply two tailored attention sub-modules on them respectively and then fuse them with different weights to detect final results. Experiments demonstrate BFAM achieves promising gains in small object detection on PASCAL VOC 2012 and Seaship datasets.