Saliency Hierarchy Modeling via Generative Kernels for Salient Object Detection

Zhang, Wenhu; Zheng, Liangli; Wang, Huanyu; Wu, Xintian; Li, Xi

doi:10.1007/978-3-031-19815-1_33

Cited by 5 publications

(1 citation statement)

References 48 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We compare with 18 state-of-the-art SOD methods, including PiCANet [22], AFNet [10], BANet [30], EG-Net [52], SCRN [40], PoolNet [20], CPD [39], BASNet [27], GateNet [53], U2Net [28], DFI [19], GCPANet [3], ITSD [55], DNA [23], PurNet [17], CTDNet [54], EDN [38] and SHNet [50] to verify the effectiveness of our method. The saliency maps published by the authors of the above methods are used and evaluated using the same validation code for a fair comparison.…”

Section: Performance Comparisonmentioning

confidence: 99%

Dual-Branch Network of Information Mutual Optimization for Salient Object Detection

2023

View full text Add to dashboard Cite

Salient object detection (SOD) is to segment significant regions of images. Noticing that the saliency maps in existing SOD methods suffer from blurring boundaries owing to insufficient extraction of boundary features and inadequate fusion between boundary features and salient region features, a dualbranch network of information mutual optimization (DIMONet) is proposed. The DIMONet has a region detection branch and a boundary detection branch to extract the corresponding features simultaneously and is mainly composed of two components. One is the mutual optimization module (MOM) that refines salient region features and boundary features based on their internal relationship. The other is the fusion module of multi-receptive fields (FMMF) that integrates multi-layer features with the refined features to distinguish salient objects better and sharpen their boundaries. With the help of MOMs and FMMFs, noises from the background in the boundary features are gradually reduced and hence the boundaries of the salient regions get sharpened. Experiments on five benchmark datasets show that our method is superior to the 18 state-ofthe-art methods.

show abstract