基于全卷积网络的语义显著性区域检测方法研究

摘要

基于底层视觉特征和先验知识的显著性区域检测算法难以检测一些复杂的显著性目标,人的视觉系统能分辨这些目标是由于其中包含丰富的语义知识.本文构建了一个基于全卷积结构的语义显著性区域检测网络,用数据驱动的方式构建从图像底层特征到人类语义认知的映射,提取语义显著性区域.针对网络提取的语义显著性区域的缺点,本文进一步引入颜色信息、目标边界信息、空间一致性信息获得准确的超像素级前景和背景概率.最后提出一个优化模型融合前景和背景概率信息、语义信息、空间一致性信息得到最终的显著性区域图.在6个数据集上与15种最新算法的比较实验证明了本文算法的有效性和鲁棒性.%The existing salient region detection algorithms based on visual stimulus and prior knowledge are difficult to detect some complicated salient regions.The human vision system can distinguish these complicated salient regions because of the rich semantic knowledge in the human visual system.We construct a semantic salient region detection network using the fully convolutional structure.Learning the mapping from the low-level features to the human semantic cognition,our network can extract semantic salient region effectively.Aiming to the defects of the semantic salient region map,we introduce the color information,object boundary information and spatial consistency information to derive accurate superpixellevel foreground and background probability.At last,we fuse the foreground and background probability,semantic information and spatial consistency information to derive the final salient region map.The experiments comparing with the state-of-the-art 15 algorithms on 6 data sets demonstrate the effectiveness of our algorithm.