Abstract
Salient object segmentation aims at distinguishing various salient objectsfrom backgrounds. Despite the lack of semantic consistency, salient objectsoften have obvious texture and location characteristics in local area. Based onthis priori, we propose a novel Local Context Attention Network (LCANet) togenerate locally reinforcement feature maps in a uniform representationalarchitecture. The proposed network introduces an Attentional Correlation Filter(ACF) module to generate explicit local attention by calculating thecorrelation feature map between coarse prediction and global context. Then itis expanded to a Local Context Block(LCB). Furthermore, an one-stagecoarse-to-fine structure is implemented based on LCB to adaptively enhance thelocal context description ability. Comprehensive experiments are conducted onseveral salient object segmentation datasets, demonstrating the superiorperformance of the proposed LCANet against the state-of-the-art methods,especially with 0.883 max F-score and 0.034 MAE on DUTS-TE dataset.