基于场景分析和视觉注意力的目标搜索

项目名称： 基于场景分析和视觉注意力的目标搜索

项目编号： No.61075018

项目类型： 面上项目

立项/批准年度： 2011

项目学科： 环境科学、安全科学

项目作者： 胡小鹏

作者单位： 大连理工大学

项目金额： 10万元

中文摘要： 本课题探索将视觉注意力选择机制与全面细致的场景分析相结合，提高复杂自然场景条件下的计算机视觉搜索能力。理论意义和实用价值在于：（1）充分的场景分析是感知复杂并且不确定自然环境的前提，学习和利用人类视觉注意力选择机制是优化分配视觉资源的有效途径，两者的紧密结合可提高计算机视觉系统在复杂自然场景条件下搜索目标的能力；（2）现有视觉注意力选择机制不能有效地解决自然场景中高复杂度和不确定性难题，他们在工程领域的应用受到了限制，因而，有必要研究新理论和新方法。本项目提出了一种基于贝叶斯推理的视觉搜索理论框架：（1）通过特征域及空间域的场景分析和对相关经验知识的利用，建立全局和局部场景表达；（2）通过特征评估及选择，提高视觉描述和注意力选择机制的效率。该框架的主要特点在于：视觉注意力选择被视为是面向搜索目标和面向场景背景两类搜索活动的综合结果，这两类搜索活动分别是基于自顶向下和自底向上的信息综合。

中文关键词： 视觉搜索；视觉注意力选择；特征评估和选择；贝叶斯推理

英文摘要： This research is to investigate the way to improve the visual search capability of computer vision in complex natural environments by integrating detailed scene analysis into a visual attention selection mechanism. The theoretical and practical value lies on the facts that (1) detailed scene analysis is necessary for understanding complex and ambiguous natural senses and, as a result, the combination of it with visual attention mechanisms provides a possibility to improve the efficiency of visual search in natural environments; (2) current computational attention selection mechanisms are not able to cope with the complexities and ambiguities of natural scenes and, accordingly, have found limited practical applications. This project presents a novel theoretical visual search framework based on Bayesian inference. In the framework, the global and local representations of the visual scene are generated through feature-space and spatial-domain scene analysis by utilizing relevant knowledge and experience. The efficiency of the visual description and attention mechanism is improved through visual feature evaluation and selection. One of the key elements of the proposed framework is that visual attention selection is regarded as the consequence of the integration of both target and background visual search activities, each of which results from the integration of both top-down and bottom-up information.

英文关键词： Visual Search; Visual Attention Selection; Feature Evalutation and Selection; Bayesian Inference