Research On Adaptive Image Re-Ranking And Greedy Selection

In order to analyze the image retrieval problem in the multimedia field, image search re-ranking has gained more and more attention in the past several years. Re-ranking is defined as constructing the retrieval results according to the initial search list via different algorithms. It aims to position the proper image to the front of the search result lists. In order to perform image re-ranking in a better way, we mainly focus on two aspects. (1) Extract better descriptive features from the image. (2) Construct a better re-ranking model. Through extracting visual features and constructing re-ranking model, we can make full use of the visual information from the images to help us re-rank the images in a better way.Due to the deficiencies existing in the image re-ranking models, two different methods are proposed as follows. (1) Query-adaptive re-ranking is applied by con-ducting query difficulty estimation for each different query. (2) The effective image greedy selection strategy is introduced to perform image retrieval for each query.(1) We apply query-adaptive re-ranking by conducting query difficulty estimation for each different query. In the feature part, deep convolutional neural network is applied to extract features from images. In the model part, we apply Visual Rank algorithm. In order to perform query-adaptive re-ranking, query difficulty estimation is proposed for each query. For different queries, huge variance exists among queries. Thus different parameters tuning strategies need to be designed for different queries. Using effective query difficulty estimation, different weighting vector and damping value is designed carefully to achieve better retrieval performance.(2) We introduce the effective greedy selection strategy to perform image retrieval for each query. At first, we need to locate the seed image, which has the ability to represent the query itself. A simple yet effective counting scheme is designed to filter the false positive image samples from the initial list. In addition, instead of conducting image re-ranking for different queries, we can focus on finding images which are similar to the seed image. Using a simple greedy strategy, we can retrieve the images from different queries effectively. We order images by their sequences of being added to the seed image set. Last but not least, multiple seed images can be applied to perform greedy selection multiple times in order to reduce the variance caused by using single seed image. Different ranks are derived from these graphs and then simple rank fusion method is applied to derive the final result in order to reduce the high variance.The main research direction of this dissertation is the completion of the above two image search re-ranking algorithms for image retrieval, and a lot of experimental results have proven the effectiveness of these algorithms. The insufficient discussion of the algorithms will be concerned in the future work.