The Study On Ranking In Online Community Image Retrieval

As the popularization of web2.0 and the development of multimedia technology, social image community emerged as a kind of network applications. The community-contributed social web service is booming and attracts lots of users to upload and share their local images online everyday. At the same time, it allows users to give tags and descriptions to the uploaded images. Trying to find more effective way for organizing, managing and retrieving abundant images has been an important issue in both academia and industry.Content-based image retrieval (CBIR) extracts low-level visual features (color, texture, shape, etc.) to index images. The main difficulty faced by CBIR is the so-called semantic gap problem, that is, the low-level visual features can not effectively express the semantic content of an image; Text-based Image retrieval depends on the text context of web images to determine the semantic content, such as title, ALT tag, anchor text, user-supplied tags and descriptions, etc. TBIR is limited by the fact that the quality of the surrounding text is not so high. With lots of noises, it is difficult to determine which textual information is truly relevant to the images. In this paper, we make full use of the textual and visual information through the effective fusing of them to solve the problems mentioned above.What's more, as the explosive growth of the number of web images, the returned search results become more and more. However, users usually just focus on few results that ranked in the front. So "good" results ranked in the front is a basic requirement of an image retrieval system. The "good" means that the returned images are relevant and at the same time diverse. In some works that only concerned relevance, each image associated with the query term is considered separately. The relationship among the results is ignored. Although search results that contain many nearly duplicated images may have a high relevance, but they provide little information to users. Moreover, query term based search manner inevitably leads to ambiguity. Different users submit a same query term may have different search intentions. In order to meet different needs as much as possible, the diversity of the search results is also very important. To this end, this paper presents a greedy visual diversity penalty algorithm to ensure both relevant and diverse results.This paper focuses on the ranking problem in online community image retrieval. The main contributions are as follows:1. Select and extract the textual and visual features of social images. Based on the textual and visual similarities among images, two graphs are constructed:textual similarity graph and visual similarity graph. Then random walks on the graphs are performed to derive two initial image ranking list.2. In order to leverage image contents and users annotations, two fusion strategies are proposed to fuse the results derived from the random walks:Fused by Ranking Score and Fused by Ranked Position. A comparison is also conducted in experiments.3. To derive both relevant and diverse set of images as top ranking results, a greedy visual diversity penalty algorithm is proposed. At each iteration of the algorithm, penalty is imposed to images using the pair-wise visual similarities. The more an image is similar to the selected one, the more penalties it receives and its ranking score is decreased. This penalty process ensures the diversity of final results.Finally, experiments are conducted to demonstrate the effectiveness of the presented model on criterions P@n, MAP, NDCG and Average Image.