基于耦合神经网络的立体匹配法

Dense Stereo Matching Method Based on PCNN

针对区域立体匹配领域窗口尺寸选择困难的问题,设计一种基于快速小窗口粗匹配和耦合神经网络细化视差图的新方法。该方法不仅能体现窗口较小时边缘定位准确的优点,也能够较为准确地恢复物体内部弱纹理区域的深度,有效地减少视差图上的黑洞和斑块。实验结果表明,该方法有很高的精确性,并且对于较小尺寸的图片有一定的实时性。

A key problem in stereo matching lies in selecting an appropriate window size, designs a new method based on using small window for first-step matching and Pulse Coupled Neural Network for perfecting disparity maps. This algorithm not only reflects the predominance that small window achieves sharper counter, but also gains accurate depth of the region with weak texture and reduces patches effectively. The experimental result indicates that this method could build dense disparity maps with high accuracy compared with common ways and there is a certain degree of real-time for smaller size of the picture.