A neural network model of the ideal observer and the statistical efficiency analysis

本研究的目的如下：首先，我们计算了一个反映人类视觉系统相对于理想观察者的可分辨性的统计效率，以便研究视觉系统的低级和高级过程的一些特性。其次，通过学习统计上最优的视觉任务响应，构建理想观测器的神经网络模型。此外，我们使用一种算法人为地扰乱网络，使其模拟人类的行为。研究结果如下：(1)计算了基于深度噪声的三维波纹表面对称图案检测任务的统计效率。利用这些结果，我们可以建立一个关于曲面构造和模式检测的模型。(2)利用统计效率的度量方法，研究了运动视差与双目视差在三维倾斜感知中的积分问题。我们明确了运动视差和双目视差信息的采样差异。(3)利用静态效率的测量方法，研究了线方向的时空总和，阐明了时间采样优于空间采样的机理。(4)构建了一种学习算法，该算法是理想观测器神经网络模型的基础。(5)利用上述算法，建立了结构-运动任务中理想观测器的双步神经网络模型。(6)我们提出了一种算法，该算法在网络中引起一些混乱，并模拟人类的行为。

The purpose of this study was as follows : First, we calculate a statistical efficiency which reflects relative discriminability of the human visual system to that of an ideal observer in order to investigate some properties in lower to higher processes of the visual system. Secondly, we construct a neural network model of the ideal observer by learning a statistically optimal responses in visual tasks. Furthermore, we artificially disorder the network with use of an algorithm and make it to simulate the human performance.We obtained the following results :(1) We calculated the statistical efficiency in the task of detecting symmetry patterns on the three dimensional (3D) corrugate surface with depth noise. With using these results, we could build a model about a surface construction and pattern detection in it.(2) We studied a problem about an integration of motion parallax with binocular disparity in the 3D slant perception with use of measurement of the statistical efficiency. We made clear the sampling differences between the motion parallax and binocular disparity information.(3) We studied a spatial and temporal summation of orientations of lines with use of measurements of the statical efficiency, and clarified the mechanism of the temporal sampling superiority over the spatial sampling.(4) We constructed an learning algorithm which was a basis of the neural network model of the ideal observer.(5) With use of the above algorithm, we built a double-steps neural network model of the ideal observer in the structure-from-motion task.(6) We made an algorithm which causes some disorder in the network and simulates a human performance.