On Combining Sequence Alignment and Feature-Quantization for Sub-Image Searching

The availability of various photo archives and photo sharing systems made similarity searching much more important because the photos are not usually conveniently tagged. So the photos (images) need to be searched by their content. Moreover, it is important not only to compare images with a query holistically but also to locate images that contain the query as their part. The query can be a picture of a person, building, or an abstract object and the task is to retrieve images of the query object but from a different perspective or images capturing a global scene containing the query object. This retrieval is called the sub-image searching. In this paper, the authors propose an algorithm, called SASISA, for retrieving database images by their similarity to and containment of a query. The novelty of it lies in application of a sequence alignment algorithm, which is commonly used in text retrieval. This forms an orthogonal solution to currently used approaches based on inverted files. They improve efficiency of SASISA by applying vector-quantization of local image feature descriptors. The proposed algorithm and its optimization are evaluated on a real-life data set containing photographs where images of logos are searched. It is compared to a state-of-the-art method (Joly & Buisson, 2009) and the improvement of 16% in mean average precision (mAP) is obtained.