Image Feature Analysis And Retrieval Algorithm Design Based On Local Binary Pattern

Faced with a flood of information in the era of big data, the retrievals of data and texts are necessary and the retrieval of multimedia files is also crucial. Digital image is one of the most important visual media. Therefore, fast and accurate image retrieval technology is urgent. In the early text-based image retrievals(TBIR), artificial text markup methods take a lot of manpower and material resources. To make matters worse, it is difficult for TBIR to accurately express image characters and contents, accordingly, content-based image retrieval(CBIR) emerges.In CBIR, the effective features of images are extracted. The similarity(or dissimilarity) between the features of testing image and database images are measured. The images in the database, which are similar to the testing image, are returned as the retrieved results.This dissertation reviews the knowledge related to digital image retrieval, the related knowledge provides the proposed schemes with the theoretical foundation. Several types of image features can be used for CBIR. Local binary patterns are gray-invariant, easy to compute and implement, so they have become widely used texture features for digital image retrieval. This paper describes the basic principles of various local binary patterns(LBP), including basic LBP, extended LBP, uniform LBP, rotation invariant LBP, rotation-invariant uniform LBP and mean LBP that was proposed as a state-of-the art technology recently. In this paper,the following work was:the features of the LBPs are analyzed, and the CBIRs are implemented based on the LBPs. The sufficient experimental results illustrate that the improved LBPs can reduce the number of data, but are probable to lead to information loss and accordingly decrease retrieve accuracy. By comparison, basic LBP and mean LBP have desired retrieval performances.Due to the strengths of basic LBP and mean LBP,their feature selection algorithms are proposed, in which the histogram frequencies are selected based on the properties of their features. The advantages of the proposed feature selection algorithms include three points: The number of non-zero features decreases so that the memory capacity is reduced. The computation operations in matching stage are reduced, which simplifies the computational complexity. The retrieval accuracy is improved. Three proposed selection criteria for feature selection algorithms of LBPs. Frequency threshold criterion: The high frequency features are selected by setting threshold. Frequency sequence criterion: A fixed number of frequency features are selected according to their frequencies from high to low values. Energy ratio criterion: Some frequency features are selected according to their frequencies from high to low values, the sum of the selected frequencies achieves the set energy ratio. Through the comprehensive experiments, this thesis amply analyzes and discusses the frequency threshold of “frequency threshold criterion”, the number of selected frequencies in “frequency sequence criterion”, and the proportion of “energy ratio criterion”. The parameters for the three criteria are optimized finally.LBP image features are analyzed and the image retrieval algorithms are designed base on three feature selection criteria, whose regularities are summarized. These conclusions are not only helpful to the performance improvement of digital image retrieval, but also provide a reference for the related research on feature selection.