Object localization and edge refinement network for salient object detection

Yao, Zhaojian; Wang, Luping

doi:10.1016/j.eswa.2022.118973

Cited by 10 publications

(1 citation statement)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We compare our model with 26 state-of-the-art SOD methods, which are BASNet [51], CPD [52], PoolNet [53], SCRN [54], EGNet [24], DFI [55], U2-Net [56], GCPANet [57], F3Net [58], GateNet [59], ITSD [60], MINet [23], LDF [61], PSGL-Net [95], Auto-MSFNet [96], VST [97], PFSNet [98], ICON [68], OLER [69] To ensure fairness, we used the evaluation code provided by https://github.com/jiwei0921/Saliency-Evaluation-Toolbox to compare the detection performance of all methods. To prove the more powerful performance and generalization ability of our method, we evaluate using five wellknown datasets, and the evaluation criteria were divided into three aspects: detection performance, efficiency performance and comprehensive performance.…”

Section: Performance Comparisonmentioning

confidence: 99%

LARNet: Towards Lightweight, Accurate and Real-Time Salient Object Detection

Wang,

Zhang,

Liu

et al. 2024

IEEE Trans. Multimedia

View full text Add to dashboard Cite

Salient object detection (SOD) has rapidly developed 1 in recent years, and detection performance has greatly improved. 2 However, the price of these improvements is increasingly complex 3 networks that require more computing resources and sacrifice 4 real-time performance. This makes it difficult to deploy these 5 approaches on devices with limited computing resources (such as 6 mobile phones, embedded platforms, etc.). Considering recently 7 developed lightweight SOD models, their detection and real-time 8 performance are always compromised in demanding practical 9 application scenarios. To solve these problems, we propose a novel 10 lightweight SOD method called LARNet and its corresponding 11 extremely lightweight method LARNet * according to application 12 requirements. These methods balance the relationship between 13 lightweight requirements, detection accuracy and real-time per-14 formance. First, we propose a saliency backbone network tailored 15 for SOD, which removes the need for pre-training with ImageNet 16 and effectively reduces feature redundancy. Subsequently, we 17 propose a novel context gating module (CGM), which simulates 18 the physiological mechanism of human brain neurons and visual 19 information processing, and realizes the deep fusion of multi-20 level features at the global level. Finally, the saliency map is 21 output after fusion of multi-level features. Extensive experiments 22 on popular benchmark datasets demonstrate that the proposed 23 LARNet (LARNet * ) achieves 98 (113) FPS on a GPU and 3 (6) 24 FPS on a CPU. With approximately 680K (90K) parameters, the 25 model has significant performance advantages over (extremely) 26 lightweight methods, even surpassing some heavyweight models.

show abstract