Robust Salient Object Detection via Adversarial Training
Yunhao Pan, Chenhong Sui, Haipeng Wang, Hao Liu, Guobin Yang, Ao Wang, Qingtao Gong · 2023
Deep salient object detection has experienced noticeable progress. Unfortunately, most existing methods focus on clean samples regardless of the noise disturbance induced by human or natural factors. This results in the detection performance being extremely vulnerable to small perturbations. To this end, this paper proposes robust salient object detection via adversarial training (ATSOD). In specific, we introduce the classical DSS algorithm and inject it into an adversarial training framework favoring salient object detection. This ensures that, apart from clean samples, adversarial examples involving tiny disturbances are also explored for model training. Comparative experiments are conducted on five popular benchmarks. Experimental results show that despite the slight performance degradation for natural examples, there is a significant performance improvement for adversarial examples.