Experimental results demonstrated the goodness of the diffusion mechanism for several computer vision tasks: image retrieval, semi-supervised and supervised learning, image classification. Diffusion requires the construction of a kNN graph in order to work. As predictable, the quality of the created graph influences the final results. Unfortunately, the larger the used dataset is, the more time the construction of the kNN graph takes, since the number of edges between nodes grows exponentially. A common and effective solution to deal with this problem is the brute-force method, but it requires a very long computation on large datasets. This paper proposes improvements on LSH kNN graph method that efficiently create an approximate kNN graph which is demonstrated to be faster than other state-of-the-art methods (18x faster than brute force on a dataset of more than 100k images) for content-based image retrieval, while obtaining also comparable performance in terms of accuracy. LSH kNN graph has been tested and compared with the state-of-the-art approaches for image retrieval on several public datasets, such as Oxford5k, \({\mathcal {R}}\)Oxford5k, Paris6k, \({\mathcal {R}}\)Paris6k and Oxford105k.

实验结果证明了扩散机制对多种计算机视觉任务的好处：图像检索，半监督和监督学习，图像分类。扩散需要构造kNN图才能起作用。可以预见，所创建图形的质量会影响最终结果。不幸的是，使用的数据集越大，由于节点之间的边数呈指数增长，因此构造kNN图花费的时间越多。解决此问题的常用方法是蛮力法，但它需要对大型数据集进行很长的计算。本文提出了对LSH kNN图方法的改进，该方法可有效地创建一个近似的kNN图，该图被证明比其他最新方法要快（在超过10万张图像的数据集上比暴力破解快18倍），其内容-基于图像的检索，同时在准确性方面也获得了可比的性能。已对LSH kNN图进行了测试，并将其与用于在多个公共数据集（例如Oxford5k，\（{\ mathcal {R}} \） Oxford5k，Paris6k，\（{\ mathcal {R}} \） Paris6k和Oxford105k。