Book retrieval based on Near-Duplicate image matching

Image Near-Duplicate (IND) detection and retrieval are both useful techniques in a variety of applications. There exist several detectors demonstrated high effectiveness and efficiency in IND detection and retrieval such as Scale Invariant Feature Transform (SIFT) detector and Histograms of Oriented Gradients (HOG) detector. We notice that any artificially taken book cover photo is a variant of the standard image about the identical book, and their relationship can be considered as IND. We propose a method for searching any book from numerous books based on IND retrieval. One of application contexts of this method is selecting a book from several similar ones. That means consumers can take photos of any book they want to have a general idea of by their cellphones to obtain some basic information by searching the corresponding book. We adopt the SIFT algorithm as our feature extraction algorithm and improve it in details to be faster and more accurate in book cover image retrieval. A two-level structure is designed in the matching stage, namely the “routine matching” and the “refining matching”. In the experiment, we establish a database of 2,000 standard image as template images and 1,200 artificially taken photos as query images. For evaluation the results, we use precision metric. The experimental results indicate that our method shows high effectiveness and efficiency.