For any node in hierarchical model, the recognition information lies in three channels：α channel recognizes current node directly based on the global features of the object; β channel computes current node by bottom - up binding of its child nodes ; γchannel predicts current node from its parent node in a top -down manner. We isolate and test the three channels separately by some specific measures, such as time control of the presentation, adjustment of image scale, occlusion, and excess pixels cropping, so that the independent information contributions from each channel can be obtained. Experiment results show that α channel is stronger than the other two channels in general, while β channel dominates the recognition task in occluded condition, the discriminative power of γ channel plays a very important role for object recognition in low resolution image. The test performance based on human vision system in this paper can become the yardstick against all computer algorithms, and meanwhile it also shreds some light on the problem of how to improve the recognition performance for computer algorithms.%对于分层图模型中的任一节点，识别信息来源于三个通道：α通道，基于物体的整体特征识别当前节点；β通道，通过孩子节点的组合以自底向上的方式来识别物体；γ通道，借助父节点以自顶向下的方式预测。通过引导受试者的注意力，控制测试数据的尺度大小和显示时间，以及遮挡和裁剪等措施来排除通道之间的相互影响，以便单独测试各通道的信息贡献量。实验结果表明：一般情况下α通道的贡献要更大；当存在遮挡时，主要依靠β通道的贡献；当分辨率很低导致本征信息不足时，γ通道将在物体识别任务中发挥主导作用。这些基于人类视觉系统的测试结果，对于相应的计算机物体识别算法性能的提高具有参考价值和指导意义。
展开▼
