SURF binarization and fast codebook construction for image retrieval

Shi-Chao Kan, Yi-Gang Cen, Yi Cen, Yanhong Wang, Viacheslav V. Voronin, Vladimir Mladenovic, Ming Zeng

Abstract: Highlights • SURF binarization and dimensionality reduction are proposed to reduce the 64-dimensional SURF descriptors into 8-dimensional descriptors. • A two-step clustering algorithm is proposed for clustering large scale samples. • A scalable overlapping partition method is proposed to reduce the computational cost for object search. Abstract A new framework for image retrieval/object search is proposed based on the VLAD model and SURF descriptors to improve the codebook construction speed, the image matching accuracy, and the online retrieval speed and to reduce the data storage. First, SURF binarization and dimensionality reduction methods are proposed to convert a 64-dimensional SURF descriptor into an 8-dimensional descriptor. Second, a two-step clustering algorithm is proposed for codebook construction to significantly reduce the computational cost of clustering while maintaining the accuracy of the clustering results. Moreover, for object search, a scalable overlapping partition method is proposed to segment an image into 65 patches with different sizes so that the object can be matched quickly and efficiently. Finally, a feature fusion strategy is employed to compensate the performance degradation caused by the information loss of our proposed dimensionality reduction method. Experiments on the Holidays and Oxford datasets demonstrate the effectiveness and efficiency of the proposed algorithms.