基于生物视觉特征和视觉心理学的视频显著性检测算法

发布时间：2018-04-05 00:35

本文选题：显著性检测　切入点：非线性简化　出处：《物理学报》2017年10期

【摘要】：提出了一种空域和时域相结合的视频显著性检测算法.对单帧图像,受视觉皮层层次化感知特性和Gestalt视觉心理学的启发,提出了一种层次化的静态显著图检测方法.在底层,通过符合生物视觉特性的特征图像(双对立颜色特征及亮度特征图像)的非线性简化模型来合成特征图像,形成多个候选显著区域;在中层,根据矩阵的最小Frobenius-范数(F-范数)性质选取竞争力最强的候选显著区域作为局部显著区域;在高层,利用Gestalt视觉心理学的核心理论,对在中层得到的局部显著区域进行整合,得到具有整体感知的空域显著图.对序列帧图像,基于运动目标在位置、运动幅度和运动方向一致性的假设,对Lucas-Kanade算法检测出的光流点进行二分类,排除噪声点的干扰,并利用光流点的运动幅度来衡量运动目标运动显著性.最后,基于人类视觉对动态信息与静态信息敏感度的差异提出了一种空域和时域显著图融合的通用模型.实验结果表明,该方法能够抑制视频背景中的噪声并且解决了运动目标稀疏等问题,能够较好地从复杂场景中检测出视频中的显著区域.
[Abstract]:A video saliency detection algorithm combining spatial domain and time domain is proposed.Based on the hierarchical perception of visual cortex and the visual psychology of Gestalt, a hierarchical static salience image detection method is proposed for a single frame image.At the bottom of the layer, feature images are synthesized by a nonlinear simplified model of feature images that conform to biological visual characteristics (double contrast color features and brightness feature images), forming multiple candidate salient regions; at the middle level,According to the minimum Frobenius-norm (F- norm) property of the matrix, the most competitive candidate significant region is selected as the local significant region, and at the high level, using the core theory of Gestalt visual psychology, the local salient region obtained in the middle layer is integrated.The spatial salience map with global perception is obtained.Based on the assumption that the moving target is consistent in position, amplitude and direction of motion, the optical flow points detected by Lucas-Kanade algorithm are classified into two categories to eliminate the interference of noise points.And the motion amplitude of the light flow point is used to measure the motion significance of the moving target.Finally, based on the sensitivity of human vision to dynamic and static information, a general model of spatial and temporal saliency fusion is proposed.The experimental results show that the proposed method can suppress the noise in the video background and solve the problem of sparse moving objects, and can detect the salient regions of the video from complex scenes.
【作者单位】：延边大学工学院计算机科学与技术系智能信息处理实验室;
【基金】：吉林省科技发展计划项目(批准号:20140101186JC)资助的课题~~
【分类号】：TP391.41

【参考文献】