视觉注意计算模型及其在多目标检测中的应用研究

Visual attention allows us to selectively process the vast amount of information with which we are confronted, prioritizing some aspects of information while ignoring others by focusing on a certain location or aspect of the visual scene. Therefore, human being can search and detect object in complex scene quickly and accurately. But it is very difficult for computer to do the same thing.So, it is naturally that many researchers devote their focuses to the simulation of the mechanisms of visual attention to solve problems in computer vision such as object detecion and recognition. .This project is conducted to detect multi-object in complex scene by simulating the mechanisms of visual attenion. Firstly, We use a hierarchical model that follows the organization of visual cortex and build an increasingly complex and invariant feature representation by pooling operation, clustering and artificial neural network. A set of features including intensity, color,orientation and motion that are position invariant, scale invariant and appearance invariant can be obtained. Different features are combined to construct a saliency map, which is a simulation of bottom-up procedure in visual cortex by the interaction between ventral pathway and dorsal pathway. Secondly, the distributions of the values of different features are learned from training samples to get task information, which is similar to the top-down procedure in visual attention. The distribution functions are used as likelihood in Baysian model to calculate posterior probability. A biased saliency map can be got by multiply the likelihood probability with the saliency map. In the biased sliency map, locations have bigger values are treated as focuses of attention (FOA). Finally, based on the elegant built visual attenion shift strategy, different FOA and their related objects are located by their saliency in the biased saliency map. .Summarily,this project can simulate the procedure of visual processing, construct visual attention model combined bottom-up information with top-down information and improve the effectiveness of multi-object detection in complex scene. Moreover, the method presented in this project can also be helpful for other research such as filtering out irrelevant information and explaining related physiologyical and physchological events.

视觉注意机制能将人的注意快速指向感兴趣信息，过滤大量无关信息，使人类能够在复杂场景中迅速准确地对目标及其行为进行检测和识别。本项目旨在模拟视觉注意机制，建立自底向上和自顶向下相结合的视觉注意计算模型，提高复杂环境中多目标检测的速度和准确率。首先自底向上提取待检测图像的空间和运动等特征，模拟视皮层双通路结构整合不同特征，减少或消除复杂环境的影响，得到特征显著图；然后针对多目标检测任务，分析和估计其特征值的分布特点，并依此选择参数估计或非参数估计来计算其分布函数，模拟自顶向下的任务信息；根据所得到的任务信息和特征显著图，利用贝叶斯方法计算特征显著图的后验概率，得到偏置特征显著图，获得注意焦点；最后建立注意焦点和检测目标间的联系，制定合理的注意转移策略，实现多目标的搜索和检测。预期本项目研究不仅能提高多目标检测的效果，而且对需要过滤无关信息的相关应用会有所帮助，也有利于解释相关生理和心理现象。

本项目通过模拟视觉信息处理和注意机制，进行了图像显著性检测和显著目标提取研究，在计划时间内完成了项目研究内容，达到了研究目标。. 项目首先进行了图像特征提取研究。人类视觉系统通过视网膜细胞获取视野中物体的不同信息，并经过从简单到复杂的处理，得到能反映物体本质的特征。对于静态图像，本项目利用图像的灰度和颜色信息，采用多尺度高斯函数滤波，得到不同尺度特征，提取尺度不变性的特征。对于视频序列，采用3D Gabor模拟和半波整流来模拟V1区简单细胞感受野的响应，提取时空结合的图像特征。. 根据所得到的特征，利用不同的方法计算图像的显著性。对于静态图像，采用DOG滤波器来计算局部对比度，对多尺度高斯滤波处理的特征提取结果进行特征和特征的方差的概率统计结果作为全局稀有度，将全局和局部对比进行线性结合，得到最终的综合显著图。对于视频系列，由于场景背景的复杂性和运动的多变性，首先对所提取的时空特征进行知觉组织，模拟神经元之间的易化和抑制作用，过滤无关或干扰信息，然后扩展Itti视觉注意计算模型，计算出空间灰度显著图，空间方向显著图，运动强度显著图和运动方向显著图四种显著图，进行线性组合得到最终的综合显著图。. 开展了从显著性图提取显著目标的研究。对于静态图像，由于所计算的显著性图效果较好，即使采用的是较简单的聚类方法，也能较好地提取显著目标。对于视频系列，由于运动的复杂性，直接由最终显著图进行显著目标提取会导致误检或漏检。本项目结合空间灰度显著图，运动强度显著图，建立表达灰度信息和运动信息的掩膜，来弥补和排除利用最终综合显著图所带来的显著目标的误检或漏检问题。. 利用简单图像和通用图像或图像库，对所提出的方法分别进行了实验分析和验证，并且与主要的方法进行结果对比，结果表明，我们所提出的方法能够取得较好的结果。本项目研究不仅能较准确地提取图像中的显著目标，为复杂背景中的多目标检测等问题提供良好的基础，而且对需要过滤无关信息的相关应用会有所帮助，同时也有利于解释相关生理和心理现象，促进对注意机制的理解。. 项目的相关研究成果以论文形式在国内核心期刊上发表一篇，另外两篇在审稿或投稿中，同时在申请发明专利一项，加速了4名研究生的培养。