PaIaNet: position-aware and identification-aware network for low-light salient object detection

Huihui Yue, Jichang Guo, Xiangjun Yin, Yi Zhang, Sida Zheng

Abstract: Due to insufficient photons and undesirable noise, salient objects in low-light scenes are ambiguous, thus limiting the performance of existing salient object detection (SOD) works. To solve this problem, inspired by the hunting mechanism of predators in biology, we propose a position-aware and identification-aware network (PaIaNet) for SOD. First, we design a position-aware decoder (PaD) for obtaining position encodes by locating the edges and main bodies of salient objects. Second, we construct an identification-aware decoder (IaD) to reason accurate saliency maps by aggregating adjacent features under the guidance of position encodes. Moreover, we propose a reverse loss to suppress background interference effectively. Extensive experiments demonstrate that our method performs favorably from comparisons of qualitative and quantitative evaluations against other state-of-the-art methods in SOD of low-light images, and even achieves competitive performance when extended to normal-light scenes. Code will be available at https://github.com/yuehuihui000/PaIaNet.