Abstract: In order to solve the current image matching algorithm which does not apply to the application of strong real-
time requirements, a parallel image matching algorithm PLS（Partial Least Squares） in combination with the law of cosines is proposed. In the CUDA architecture, the algorithm puts the image matrix into blocks. Each small image block is stored in the shared memory and the image features of each piece are extracted, and then the image features which are combined use the law of cosines to calculate similarity, and the matching image is found. The experiments show that the algorithm can achieve the parallel image matching in the CUDA architecture, compared to the CPU serial image matching. The timeliness of the proposed algorithm is hundreds of times compared to the CPU serial algorithm.

Key words: Compute Unified Device Architecture（CUDA）, GPU technology, Partial Least Squares（PLS） method, parallel computing, law of cosines, image matching

摘要： 为解决目前已有的图像匹配算法不适用于对实时性要求很强的应用，提出了PLS（Partial Least Squares）与余弦定理相结合的并行化图像匹配算法。该算法在CUDA架构下，对图像矩阵分块，分块后每个小块图像存入共享存储器处理并提取每个小块图像特征，通过合并后图像特征采用余弦定理计算图像的相似度，从而找出匹配图像。实验表明，CUDA架构下可以实现图像的并行匹配，与CPU上串行匹配相比，时效性提高了百倍以上。

关键词: 统一设备计算架构（CUDA）, GPU技术, 偏最小二乘（PLS）方法, 并行计算, 余弦定理, 图像匹配