媒体场景视觉显著计算

Scene Salience Computation for Media

借鉴人视觉系统中一体化的多种显著控制策略,本文建立视频显著内容理解的计算模型.计算模型由控制空间模型和面向实现的网格模型组成.控制空间整合多个控制视频内容理解方式的因素,网格将控制空间的控制策略映射到具体实现技术.接着,本文提出一种遵照以上计算模型的多通路场景显著计算方法.该方法利用多尺度金字塔分解及中心-环绕差计算得到亮度、颜色和方向3个静态分显著图,采用背景注册技术抽取场景中动态信息及亮度中心环绕差运算得到场景正规化的动态分显著图,用具有自适应链接强度的脉冲耦合神经网络融合成最终的显著图.对比实验表明本文方法能兼顾场景中动静显著信息,在一定程度上实现视觉局部对比、关键特征、场景全局长时特性、场景语义联系、当前任务等各个因素统一作用控制支配内容显著理解.

Referring to integrating attention muti-control strategies in human visual system, computational model for video salience a- nalysis are built in this paper. The model is composed of control space model and grid model facing implement. In control space, multi-control strategies for video content understanding are unified. And, control strategy is mapped into implementation by grid. Then,a multi-path conspicuity computing method （MCCM） conforming to computation model for video scene analysis is presented. First, multi resolution pyramids and centre-surround difference through across-scale difference are used in order to compute intensity feature maps, color feature maps, and orientation feature maps for conspicuity. Then, dynamic information areas are extracted with background registration technique. After that, dynamic feature conspicuity map is obtained following intensity centre-surround differ- ence and normalization operator. Finally, four feature conspicuity maps are fused using Pulse Coupled Neural network （ PCNN） with adaptive linking strength parameter. Experiment results contrast to other method show that the method in this paper can compute whole conspicuity maps with dynamic and still saliency information, which is suited for media scene analysis.