Evaluation On The Impact Of Image Quality On Image Retrieval

In recent years, a content-based image retrieval (CBIR) using local invariant features has been a hot research topic. CBIR system always adopts local invariant features. In CBIR, the retrieval performance, such as accuracy and efficiency, is significantly affected by the quality of the query image. Generally, the image quality is determined by many factors, such as image resolution, noise addition, rotation, JPEG compression, selected local features, etc. In this thesis, we make a comprehensive study on those factors to investigate their impact on image search accuracy. In this thesis, the main work and innovations are as follows:1. The overview of the steps of content-based image retrieval, including part of online and offline. The part of offline includes image feature extraction, training of code book, feature quantization, image indexing. Online part includes image feature extraction, feature quantization, indexing and searching.2. Building the baseline system with the classic Bag-of-Visual-Words model and the inverted index structure. Two public released datasets, i.e., UKBench and Oxford Building, are selected as ground truth dataset. Based on the extensive experimental study, some conclusions are drawn from the evaluation results. In UKBench dataset, performance keeps stable if the image size is controlled in a certain range of 384 times 288 to 576 times 432. Also, the JPEG compression ratio can be reduced to as low as 8% of base one but has little impact on retrieval performance. What is more, Image performance achieves 90% of best result though its PSNR value is 32 when we test in Oxford Building dataset. Experiments show that reducing the quality of query images in a reasonable rang does not harm the retrieval result. What’s more, data produced from this thesis benefit application based on content-based image retrieval.