基于全局感知和局部细化的夜间显著目标检测

Salient object detection in nighttime images via global awareness and local refinement

为有效解决由于信噪比偏低,夜间图像显著目标检测结果具有区域定位不准确、内部结构不完整和外部边界不清晰等问题,提出由全局语义感知和局部结构细化引导的深度全卷积网络用于夜间显著性检测,判别夜间图像的像素级显著性。为进一步完善目标定位和像素分类性能,模型集成全局语义感知模块以获取更多空间信息,嵌入局部结构细化模块提取具有完整边界知识的显著性信息。实验结果表明,所提模型具备先进的夜间显著性检测性能。

Due to the low signal-to-noise ratio, the results of salient object detection in night images are not accurate, the internal structure is incomplete and the external boundary is not clear, a deep full convolutional network guided by global semantic awareness(GSA) and local structure refinement(LSR) for nighttime saliency detection to discriminate the pixel-level saliency of nighttime images was proposed. To further improve the object location and pixel classification performance, the proposed model integrated global semantic awareness modules to preserve more spatial information, and local structure refinement modules were embedded to extract saliency information with complete boundary knowledge. Experimental results show that the proposed model has advanced nighttime saliency detection performance.