基于协作Hopfield网络的迭代立体匹配算法

针对立体匹配算法中求解能量函数全局最小问题,提出一种基于协作Hopfield网络的迭代立体匹配算法.它采用两个具有相似结构的Hopfield神经网络协作求解匹配问题,两个网络的不同之处是匹配过程中所采用的基准图不同.然后根据左右一致性约束实现两个Hopfield网络之间的协作,从而避免落入局部最小.为加快收敛速度,该算法将视差图的最优搜索问题转换为二值神经网络的迭代收敛过程.利用局部匹配算法的结果预标记初始视差,以设定神经网络初始权重.并根据局部匹配算法中隐含的假定条件,提出了局部匹配算法视差结果的评估准则,以确定各像素的视差搜索范围,从而减少各次迭代过程中状态待确定的神经元个数.实验表明该方法在性能和收敛速度上都要优于传统的Boltzmann机方法.

An Iterative Approach to Stereo Matching Based on Cooperative Hopfield Networks

In order to solve the energy function minimization in stereo matching, an iterative approach based on cooperative Hopfield networks is proposed. This approach uses two Hopfield networks, with similar structure, to solve the matching problem cooperatively. According to the mutual correspondence constraint, a cooperation strategy between two Hopfield networks is presented to avoid the algorithm falling into local minima early. To shorten the convergence time, the optimal search problem of disparity map is converted to an iterative convergence process of bi-valued neural networks. The disparity pre-labeling based on local matching is used to initialize the weights of the neural networks. Then according to the implicit assumption in the local matching algorithm, two evaluation criteria are applied to determine the disparity range of each pixel for reducing the number of neurons with uncertain status in each iteration. Experiments indicate this approach is much better than Boltzmann machine method in performance and convergence speed.