项目名称: 基于深度学习的层次化视觉注意模型研究
项目编号: No.61202328
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 计算机科学学科
项目作者: 王威
作者单位: 中国科学院自动化研究所
项目金额: 24万元
中文摘要: 视觉注意建模是计算机科学、心理生理学、计算神经科学等领域的多学科交叉研究问题,目前的视觉注意模型主要模拟初级视皮层简单细胞的"中心-周围"机制,而大量的实验表明在不同层次视皮层中都存在注意机制,而且对不同的视觉特征表现出不同的选择性。本项目将研究一种基于深度学习的层次化视觉注意计算模型,这种层次化的深度学习框架不仅模拟了各级视皮层的选择性注意机制,而且学习了在自下而上数据驱动与自上而下任务驱动共同作用下的多层视觉特征、不同视觉特征在视觉注意中的重要性。一方面,本项目提出的层次化注意模型将提高对人眼关注位置的预测精度,具有很大的潜在应用价值;另一方面,在深度学习理论中引入注意机制,有望提高深度学习算法在其它视觉任务中的性能(如提高目标检测和目标识别的精度)。本研究项目具有很强的理论性和实用性。
中文关键词: 视觉注意;深度学习;深度神经网络;;
英文摘要: Modelling visual attention is an interdiscipline research topic which invloves in computer science, psychophysiology and computational neuroscience. Current visual attention models mainly focus on modelling the center-surround mechanism of the simple cells in primary visual cortex. However, large amounts of experiments show that selective attention mechanism exists in each layer of visual cortex, and different visual features play different roles in selective attention. In this project, we will propose a deep learning-based hierarchical visual attention model which not only simulates the selective attention of each visual layer with its internal driving mechanism, but also learns multi-layer visual features by combining bottom-up data and top-down task, and learns the importances of the features in each layer. For one thing, the proposed model will improve the accuracy of fixation prediction, which has a large amount of potential applications; for another, embedding attention mechanism into deep learning framework will improve its performances in several other visual tasks, such as object detection and object recognition. This project is highly theoretical and of high practicality.
英文关键词: Visual Attention;Deep Learning;Deep Neural Networks;;