视差计算的层级模型

生物的立体感知是由一个层级网络完成的:从初级视皮层到高层区域,神经元感受野逐步增大,局部立体感知逐步变为全局立体感知.视皮层中存在大量对视差敏感的神经元,其中V1区单个神经元的视差选择特性可用视差能量模型来描述.文中从生物的立体感知过程出发,提出了一种计算图像视差的层级模型,主要贡献有:(1)提出了一种符合心理学实验结果的归一化视差能量模型,减弱了图像对比度变化对神经元视差响应能量的影响;(2)利用视皮层视差功能柱的性质,提出了一种不同倾向视差神经元的汇聚方法;(3)根据不同脑皮层之间的连接关系,提出了一种两层网络结构来解决V1区神经元编码视差的歧义问题.文中方法可以有效提高纹理重复和纹理不丰富区域的视差计算精度.

A hierarchical model for disparity computation

Biological studies reveal that stereopsis is completed by a hierarchical network starting from V1. The local stereopsis gradually converges to a global one with the increasing sizes of receptive fields. Many disparity sensitive neurons exist in the cortex and the Disparity Energy Model （DEM） can describe the disparity tuning responses of these neurons in V1. Inspired by the disparity computation in the visual cortex, a hierarchical model is proposed in this paper. The merit of the proposed model lies in three aspects： （1） Consistent with psychological results, a normalized disparity energy model is proposed to weaken the influence of stereograms contrast on disparity energy of neurons; （2） based on properties of the disparity columns in the visual cortex, a new disparity pooling scheme is proposed; （3） a two-layer interacting network is constructed to solve the ambiguities of the neuron responses in V1. The resultant disparities of our model are accurate, especially in area with repetitive or low texture.