当前位置: X-MOL 学术Appl. Soft Comput. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Visual fixation prediction with incomplete attention map based on brain storm optimization
Applied Soft Computing ( IF 8.7 ) Pub Date : 2020-08-26 , DOI: 10.1016/j.asoc.2020.106653
Jian Yang , Yang Shen , Yuhui Shi

We cannot see everything around us. Instead, the visual attention mechanism will select some fixations from extensive visual information for further processing. Many computational attention models have been proposed by imitating this mechanism. However, almost all of the state-of-the-art computational models output a complete saliency map, which means one needs to go through all the regions in a scene before figuring out which part is more salient. According to the findings of neuroscience researches, it is unnecessary to evaluate every part at the glance stage of visual perception. Many researchers believe that the attention maps in different parts of our brain should be an incomplete one. What illustrated in this paper is a visual fixation prediction model that calculates the attention regions in a partially random manner. The output saliency map indeed is incomplete. We first translate the fixation prediction problem to a 2-D searching problem, then apply a newly proposed swarm intelligence algorithm known as Brain Storm Optimization (BSO) to search the fixation in different scenes. The proposed method can guide the searching process, converging to relatively more salient positions during iterations without going through all parts of an image. We evaluate the proposed method on a large scale CAT2000 dataset and the Extended Complex Scene Saliency Dataset (ECSSD). By comparatively studying with the other eight fixation prediction models and 24 salient region detection models, results indicate that the proposed method is effective in predicting the fixation rapidly, which makes it a good candidate for computational visual attention modeling.



中文翻译:

基于头脑风暴优化的不完整注意力图视觉注视预测

我们看不到周围的一切。相反,视觉注意机制将从广泛的视觉信息中选择一些注视点,以进行进一步处理。通过模仿这种机制,已经提出了许多计算注意模型。但是,几乎所有最新的计算模型都输出完整的显着图,这意味着在确定哪个部分更显着之前,需要遍历场景中的所有区域。根据神经科学研究的发现,在视觉感知的一眼阶段就不必评估每个部分。许多研究人员认为,我们大脑不同部位的注意力图应该是不完整的。本文说明的是一种视觉注视预测模型,该模型以部分随机的方式计算关注区域。输出显着性图确实不完整。我们首先将注视预测问题转换为二维搜索问题,然后应用新提出的群智能算法(称为脑风暴优化(BSO))在不同场景中搜索注视。所提出的方法可以指导搜索过程,在迭代过程中收敛到相对更显着的位置,而无需遍历图像的所有部分。我们在大型CAT2000数据集和扩展复杂场景显着性数据集(ECSSD)上评估了所提出的方法。通过与其他8个注视预测模型和24个显着区域检测模型的比较研究,结果表明,该方法可快速有效地预测注视,这使其成为计算视觉注意力建模的理想选择。我们首先将注视预测问题转换为二维搜索问题,然后应用新提出的群智能算法(称为脑风暴优化(BSO))在不同场景中搜索注视。所提出的方法可以指导搜索过程,在迭代过程中收敛到相对更多的显着位置,而无需遍历图像的所有部分。我们在大型CAT2000数据集和扩展复杂场景显着性数据集(ECSSD)上评估了所提出的方法。通过与其他8个注视预测模型和24个显着区域检测模型的比较研究,结果表明,该方法可快速有效地预测注视,这使其成为计算视觉注意力建模的理想选择。我们首先将注视预测问题转换为二维搜索问题,然后应用新提出的群智能算法(称为脑风暴优化(BSO))在不同场景中搜索注视。所提出的方法可以指导搜索过程,在迭代过程中收敛到相对更多的显着位置,而无需遍历图像的所有部分。我们在大型CAT2000数据集和扩展的复杂场景显着性数据集(ECSSD)上评估了所提出的方法。通过与其他8个注视预测模型和24个显着区域检测模型的比较研究,结果表明,该方法可快速有效地预测注视,这使其成为计算视觉注意力建模的理想选择。

更新日期:2020-08-26
down
wechat
bug