针对SIFT描述子实时性差和传统二进制描述子对尺度、旋转和视角变化鲁棒性差的问题, 本文通过优化采样模式和添加灰度差分不变量比较测试进行改进, 提出了一种鲁棒性更高的二进制描述子。首先, 设计了一种尺度关联、编号标记的采样模式; 然后, 旋转采样模式中各采样点到特定位置, 确保描述子尺度、旋转不变性; 接着, 分析了采样点点对模式对描述子的影响, 选择使用机器学习训练后的128对采样点对; 最后, 选择灰度值比较测试及梯度绝对值和比较测试构建二进制描述子。实验中采用DoG检测图像关键点, 结果表明: 本文提出的描述子在描述子构建和描述子匹配上比SIFT描述子分别快84%和67%; 在有视角变化的图像匹配上, 准确率比传统的二进制描述子高3%~5%, 召回率平均要高30%以上。本文提出的特征点描述方法适用于时间要求高的图像匹配领域。

The SIFT descriptor has poor real-time performance and conventional binary descriptors are poorly robust to scale, rotation, and viewpoint changes. They can be improved by optimizing the sampling pattern and adding gray-value differential invariant comparisons. This study proposed a binary descriptor with a higher robustness. First, a sampling pattern with scale association and number marking was presented. Then, each sampling point of the sampling pattern was rotated to a specific position to ensure that the descriptor was invariant to scale and rotation. Subsequently, the influence of sampling pairs on the descriptor was analyzed, and 128 sampling pairs after machine learning were chosen. Finally, intensity comparison and gradient absolute value comparison were selected to build the descriptor. The image keypoint detection was based on the difference of Gaussians method. Experiment results show that the proposed descriptor is 84% and 67% faster than the SIFT descriptor in descriptor construction and descriptor matching, respectively. Its accuracy is 3% to 5% higher than that of the conventional binary descriptor in image matching with view change, and the recall rate is more than 30%. The descriptor presented in this study is suitable for time-critical image matching.