摘要:人类能够迅速地选取视野中的关键部分，选择性地将视觉处理资源分配给这些视觉显著的区域.在计算机视觉领域，理解和模拟人类视觉系统的这种注意力机制，得到了学界的大力关注，并显示出了广阔的应用前景.近年来，随着计算能力的增强以及大规模显著性检测数据集的建立，深度学习技术逐渐成为视觉注意力机制计算和建模的主要手段.综述了视觉注意力检测的最新研究进展，包括人眼关注点检测和显著物体检测，并讨论了当前流行的视觉显著性检测数据集和常用的评估指标.对基于深度学习的工作进行了综述，也对之前代表性的非深度学习模型进行了讨论，同时，对这些模型在不同的数据集上的性能进行了详细评估.最后探讨了该领域的研究趋势和未来的发展方向.



Abstract:Humans have ability to quickly select a subset of the visual input and allocate processing resources to those visually important regions. In computer vision community, understanding and emulating such attention mechanism of the human visual system has attracted much attention from the researchers and shown a wide range of applications. More recently, with the ever increasing computational power and availability of large-scale saliency datasets, deep learning has become a popular tool for modeling visual attention. This review includes the recent advances in visual attention modeling, including fixation prediction and salient object detection. It also discusses popular visual attention benchmarks and various evaluation metrics. The emphasis of this review is both on the deep learning based studies and the represented non-deep learning models. Extensive experiments are also performed on various benchmarks for evaluating the performance of those visual attention models. In the end, the review highlights current research trends and provides insight into the future direction.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5636