Chen Zuyao ; Xu Qianqian ; Cong Runmin ; Huang Qingming

Abstract

Deep convolutional neural networks have achieved competitive performance insalient object detection, in which how to learn effective and comprehensivefeatures plays a critical role. Most of the previous works mainly adoptedmultiple level feature integration yet ignored the gap between differentfeatures. Besides, there also exists a dilution process of high-level featuresas they passed on the top-down pathway. To remedy these issues, we propose anovel network named GCPANet to effectively integrate low-level appearancefeatures, high-level semantic features, and global context features throughsome progressive context-aware Feature Interweaved Aggregation (FIA) modulesand generate the saliency map in a supervised way. Moreover, a Head Attention(HA) module is used to reduce information redundancy and enhance the top layersfeatures by leveraging the spatial and channel-wise attention, and the SelfRefinement (SR) module is utilized to further refine and heighten the inputfeatures. Furthermore, we design the Global Context Flow (GCF) module togenerate the global context information at different stages, which aims tolearn the relationship among different salient regions and alleviate thedilution effect of high-level features. Experimental results on six benchmarkdatasets demonstrate that the proposed approach outperforms thestate-of-the-art methods both quantitatively and qualitatively.