Deep network for visual saliency prediction by encoding image composition.
J. Visual Communication and Image Representation(2018)
摘要
This article will be visual significance into the graphical guidance (the chart is a medium-sized join subgraph) Deep structure, from the level of learning a significant map The original image pixel to the object level graphic (oGL), and further Space level graphics (sGL). In particular, we first sample Super pixels from each image, and they are used as buildings Block of each object. In order to seamlessly describe different objects The number of oGLs is generated by spatial adjacent links The super pixel oGL object response mapping is obtained by obtaining Transfer the semantics of the image tag to oGL. As space The layout of the object plays an important role in the prominence of the object Based on the relevant learning distribution proposed sGL OGL position between. Finally, in order to imitate the” winner of all” Biological vision mechanism, the largest majority of voting programs The sGL of the image is probabilistically combined into a significant graph. Experimental results show that oGLs and sGLs capture the object level well And space-level visual cues, resulting in competitiveness Significant detection accuracy.
更多查看译文
关键词
Deep network, Visual saliency, Image composition
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络