The Quality Assessment Of Image Retrieval Results

Query performance prediction is a very important research topic in image retrieval, which attempts to automatically estimate the performance of the search results returned for a given query without relevance judgments or user feedback. This technique will provide both users and image retrieval systems with a better service. Through the in-depth investigation on query performance prediction methods in text search and image retrieval, this dissertation proposes two novel query performance prediction algorithms for image retrieval:(1) one algorithm automatically estimates the average precision of the image search results;(2) the other investigates the relationship between the textual query and the returned images to predict the query performance.(1) Estimating the average precision of the image search results:the ground-truth performance of each query in image search is always measured based on manual relevance labels via the commonly used average precision (AP), so an intuitive idea for the quality assessment of the image search results is to directly measure the average precision of the images returned to a given query. According to the analysis and derivation of the AP calculation formula, this algorithm designs two different methods to estimate the average precision:a) label an image which is positively correlated with the given query as relevance and others as irrelevance; b) estimate each image’s relevance probability to the given query. With each image’s relevance to the query is estimated, we can deduce the AP of the image search results for each query.(2) Investigating the relationship between the textual query and the returned images:in general, given the images returned for an unknown query, if this query is "easy" with lots of relevant images returned, we can easily deduce the unknown query from the returned images; otherwise, if the query is "difficult" and the search results contain many irrelevant images, it is then hard to recover the original query from the search results. Based on the above observation, we can draw a conclusion that: if the visual theme of the search result is highly consistent with the input query, the search result has high quality, corresponding to an "easy" query; otherwise, the search result is poor. Therefore, the relationship between the query and the returned images plays a crucial role in query performance prediction and should be well explored.The main research direction of this dissertation is the completion of the above two query performance prediction 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.