提出一种针对嵌入式系统的图像检索算法,通过提取目标局部特征来进行图像检索。为了提高检索的实时性并兼顾正确率,选用经典SIFT特征为基础进行改进。在关键点检测阶段使用均值滤波代替高斯滤波大大提高特征提取速度。在描述符生成阶段通过稀疏矩阵将SIFT特征映射为二进制描述符。引入基于K-means的Multi-probe LSH方法对二进制描述符进行快速检索和匹配。通过一系列的图像缩放、旋转、模糊和光照变化对比实验,可以看出该算法与现有的经典算法相比在检索正确率及实时性方面均有很好的表现。%In this paper it introduces an algorithm of image retrieval for embedded system, which uses local features to do image retrieval. In order to reduce the time cost and get high precision, it improves SIFT feature and descriptor. It replaces Gauss filter with mean filter in detection of scale-space extrema stage. It projects SIFT feature into binary descriptor with sparse matrix. It searches and matches object with multi-probe LSH based on K-means. By doing a series of experiments scale, rotation, blur, illumination, it can draw a conclusion that the algorithm has better performance than traditional state of arts.
展开▼