Research On Hierachical Image Retrieval Model

In the high-pace modern society, large scale image retrieval is becoming a challenging topic, as the increasingly need for vivid and efficient information has lead human into the so-called era of picture reading. In this paper a serial of problems in current image retrieval are discussed, and the work in smoothing algorithm, feature extraction and model of image retrieval system is proposed.The paper starts with an acceleration scenarios of pyramid smoothing algorithm that decorrelate smoothing radius and calculation time to maintain relatively high real-time performance with sufficient smoothing effect. The overall performance make it preferable for image pre-processing as an alternative to conventional noise reduction algorithm and image filtering as an option, especially for cases with limited computational resources and large smooth radius.Then it gives the dynamic local color histograms (DLCH) based retrieval algorithm that synthetically considers the statistical characteristics of color and shape. The proposed technique can offer more flexibility with tunable parameters emphasize on particular characteristic descriptions.And finally a new image retrieval strategy is introduced which utilizes the proposed techniques, the hierarchical image retrieval model. The model adopts a classification-then-indexing scenario that removes unrelated items layer by layer. Hence make it possible to ensure retrieval accuracy while significantly improve real-time performance. The model can be considered as a brand new attempt to the real-time large-scale data retrieval issue.The proposed approach of hierarchical image retrieval model is evaluated on a product-based image database to demonstrate the employment and performance. In the setup, we made a detailed introduction to algorithms and involved parameters in image pre-processing, edge detection scheme and feature extraction schemes. And adaptively tuned and improved DLCH, CEDD, and LBP algorithm. The evaluation in retrieval results and run time to each algorithm indicates the hierarchical retrieval model greatly helps the system to achieve optimal performance.