如何通过立体匹配提高图像景深精度一直是机器视觉研究领域的一个难点, 针对自适应窗口算法易受光照不均和窗口形状难以有效描述待匹配图像边界等问题, 提出一种异形自适应窗口局部立体匹配算法. 由于传统 Census 变换易受中心像素波动影响, 因此提出一种像素信息三维化处理技术, 并通过窗口内非中心像素间差异和窗口间中心像素的差异信息计算匹配代价; 为更好地贴合图像轮廓, 提出了由双螺旋路径法确定的异形窗口进行代价聚合, 对像素区域同时沿两条搜索路径自适应确定形状大小, 形成比传统技术更加高效多变的匹配窗口. 实验结果表明, 相比 Middlebury 平台上其他算法, 所提算法具有更为准确的图像边界描述能力, 可有效提高立体匹配精度, 同时对于光照不均的情况具有更强的鲁棒性.

Improving the accuracy of image field depth using a stereo matching algorithm is a key challenge in the field of machine vision. A so-called shaped adaptive window local stereo matching algorithm is proposed to solve the problem that traditional adaptive window algorithms are susceptible to uneven illumination, which affects their accuracy, and that the shape of the window does not accurately describe the boundary of the image that is to be matched. Because the traditional Census transform is vulnerable to fluctuations in illumination of the central pixel, a so-called three-dimensional pixel information technique which combines with the Census transform is proposed. The matching cost is obtained by comparing the central pixel with non-central pixels and by comparing the non-central pixels with one another. Moreover, for better conformance to the image contour, and to better approximate the image boundary of the examined object to obtain higher matching accuracy by the subsequent cost aggregation operation, a double helix path method is proposed. This method obtains a window that is shaped differently from the traditional rectangle or cross-based windows. In this method, the double paths which search and examine the pixel region simultaneously shorten the time required for window construction; meanwhile, the double helix path creates a more variable window shape, which more easily conforms to the image contour. The experimental results show that, compared with most conventional stereo matching algorithms on the Middlebury platform, the proposed algorithm is better able to characterise the image boundary to improve matching precision and is more robust in cases of uneven illumination.