基于感知哈希与尺度不变特征变换的快速拼接算法

Fast Stitching Algorithm Based on Perceptual HASH and Scale Invariant Feature Transformation

摘要:

鉴于尺度不变特征变换(SIFT)结构复杂, k-d树匹配算法对于高维特征计算量过大, 对SIFT特征信息利用少并且匹配的结果有大量误差, 因此提出一种基于感知哈希与尺度不变特征变换的快速拼接算法.首先使用感知哈希算法, 提取匹配图像与待匹配图像的HASH指纹, 快速识别出两幅图像的相似部分.然后计算并提取出相似区域SIFT特征点.在特征点匹配算法上, 替换传统的k-d树算法, 利用SIFT特征点的主方向以及坐标位置信息过滤掉不必要的特征点匹配, 减少匹配耗时.最后用加权最佳拼接缝图像融合算法消除突变, 完成拼接.实验结果显示, 本文算法提取的特征点数比传统算法更少, 在匹配算法上减少计算量, 同时还粗过滤了一部分误匹配, 提高了匹配准确度, 算法的耗时较传统方法有明显提升.

Abstract:

The k-d tree registration algorithm has the disadvantages of too much computation for high-dimensional features, less use of SIFT feature information and a lot of errors in the registration results, due to the complexity of scale invariant feature transformation (SIFT). Therefore, a fast stitching algorithm based on perceptual hash and scale invariant feature transformation is proposed. Firstly, the perceptual HASH algorithm is used to extract the hash fingerprint of the matching image and that of the image to be matched, and the similar parts of the two images are quickly identified. Then the SIFT feature points of similar areas are calculated and extracted. In the feature point registration algorithm, the traditional k-d tree algorithm is replaced. The main direction and coordinate position information of SIFT feature points are used to filter out the unnecessary feature point matching and reduce the registration time. Finally, the best weighted seam image fusion algorithm is used to eliminate the mutation and complete the stitching. Experimental results show that the number of feature points extracted by this algorithm is less than the number of feature points extracted by the traditional algorithm, and the amount of calculation is reduced in the registration algorithm. At the same time, some mismatches are roughly filtered, which improves the matching accuracy, and the time-consuming of the algorithm is significantly improved compared with the traditional method.