基于视觉词典和位置敏感哈希的图像检索方法

摘要

当前视觉词袋（Bag of Visual Word，BoVW）模型中的视觉词典均由 k-means 及其改进算法在原始局部特征描述子上聚类生成，但随着图像数据的迅速增长，在原始局部特征空间中进行聚类存在着运行时间较长和占用内存较大的问题。针对着这些问题，提出了一种基于视觉词典和位置敏感哈希的图像检索方法。首先，选择合适的生成二进制哈希码的哈希算法，将局部特征点保持相似性地映射为二进制哈希码。然后，在二进制哈希码上进行 k-means，生成视觉词为二进制码的视觉词典。最后，用视觉单词的词频向量表示图像内容，根据词频向量对图像进行检索。在 SIFT-1M 和 Caltech-256数据集上的实验结果表明，本方法可以缩短视觉词典生成的时间，占用更少的存储空间，与传统的基于 k-means 的视觉词典算法相比，图像检索性能基本不变。%Currently,visual vocabulary in the bag of visual word(BoVW)model is often generated either by the k-means algorithm or its improved algorithms based on original local features,but with the increasing amount of image data,clustering in the original local feature space have the problem of long running time and large memory occupation. For these problems,we propose an image retrieval method based on visual vocabulary and locality sensitive hashing. Firstly,we select an appropriate binary hashing algorithm,which map the similar local feature points onto similar binary hash codes. Second,we play k-means on the binary hash codes and generate the visual vocabulary with binary visual word. Finally,the image is represented with the word frequency vector of visual words,and image retrieval is achieved based on the word frequency vector. Experimental results on SIFT-1M and Caltech-256 data sets show that the pro-posed method can reduce the visual vocabulary generation time and take up less storage space. Compared with the traditional image re-trieval based on k-means,the image retrieval performance is almost the same.