The Research On Web-Image Similarity

As the Internet and multimedia develop rapidly in the recent years, a huge number of digital images are widespread on the web every day. These images information make people's live information convenient, but also poses a problem. How to quickly and efficiently retrieve the images from large multimedia database to satisfy users' need? The research on web-image similarity as the key issue of image retrieval, not only has scientific value but also has practical value.This thesis focuses on the study of web-image. The issues about image representation based on image feature, image similarity measurement, and image re-ranking are carried out. The main works of this thesis are summarized as follows.Taking the low resolution, redundant information and transform complexity of web image into account, an improved approach to web-image retrieval based on non-negative sparse coding and multi-codebook is proposed. First, distinct multi-codebooks are learned from the training features. Multi-codebooks consider the spatial information of visual features to a certain extent. Second, features are encoded through non-negative sparse coding algorithm. This encoding scheme can achieve a much lower reconstruction error and capture salient properties of images. Third, an improved intersection kernel function is designed, which is time-real, stable and efficient to compute image similarity. The proposed scheme takes into account the characteristics of web-image, at the same time can represent the features of image accurately.Considering the computation complexity of state-of-the-art re-ranking approaches, an image re-ranking method with the k-nearest neighbors(K-NN) of the query is present. Database images are re-ranked according to collaboratively the returned results of query and the top-k retrieved images. As a novel rank-order based approach, it has low computation complexity and can successfully retrieve the images with large variations.There is no ready-made web-image database to do experiment. So web multiclass 6K database is collected. Experiments are carried out on 6K database and University of Kentucky Benchmark(UKB) dataset that is public internationally. The experimental results on the two databases demonstrate the proposed schemes are effective and efficiency in improving the accuracy precision of similarly image, and take into account the characteristics of this particular area of web-image. A simple demonstration platform is given to show the algorithm intuitively.