摘要
视觉注意力建模技术是计算机主动视觉领域的关键技术。为了提高传统视觉注意力模型预测人眼注视点的精度,提出了一种基于有监督学习的视觉注意力模型,该模型通过引入真实眼动数据作为监督值,结合自底向上特征(如方向、颜色、强度)及自顶向下特征(如人、脸、汽车等),作为Ada Boost分类器的样本,训练出自然场景中像素点特征向量与注视点之间的映射,并采用训练后的模型生成自然场景的视觉显著区域。实验结果表明,本文模型优于现有的8个主流模型,并能够较高质量地预测人类注视点。
One key technique of the active sense of computer vision is visual attention modeling. In order to improve the accuracy of the visual attention model in predicting eye fixations in free-viewing of natural scenes,a visual attention model based on supervised learning was proposed. Here,we introduced eye fixa-tions as supervision,and combined best bottom-up features such as orientation,color,intensity,with topdown cognitive visual features( e. g.,faces,people,cars,etc.) as sample input to train Ada Boost classifiers. Then,we learned a direct mapping form the features to eye fixations and used it to detect the most salient region in a natural scene. Experimental results show that the proposed algorithm outperforms the eight state-of-the-art models,and can predict eye fixations accurately.
出处
《中国体视学与图像分析》
2015年第3期201-207,共7页
Chinese Journal of Stereology and Image Analysis
基金
国家自然科学基金(61201319)
西北工业大学"翱翔之星"
"新人新方向"与"科研平台发展项目"资助
