Font Size: a A A

Global Contrast Of Superpixels Based Salient Region Detection

Posted on:2014-01-27Degree:MasterType:Thesis
Country:ChinaCandidate:J WangFull Text:PDF
GTID:2248330398461469Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Saliency analysis model can simulate the biological visual system and extract the significant portion of the input image, which is the most interesting and most representative of the image content. As the basic processing step of many computer vision problems, it has great research meaning and broad application prospects. Saliency analysis model, with high efficiency and quality, can abstract reliable and complete data for core region locking and tracking, object analysis, scene description, object-sensitive image retrieval and editing; it can allocate computational resources of the image analysis and processing reasonably and properly; what is more, it can deepen our awareness and understanding of the human visual system in turn.Depending on the implementation mechanism, current image significance analysis models can be broadly divided into three categories:the class of biological visual simulation, the class of pure math computation and the integration of two classes. This paper selects10iconic models for research, which cover three categories above and the latest research findings as well, and are all referenced multiple times. This paper also provides the general solution of salient region detection, consisting of three steps, screening features, calculating the saliency of the basic unit and the post-processing. It also gives a general summary of effective processing methods of each step. Based on the above study, it can be found that saliency detection is still a hot issue to be resolved, and it is necessary to do further exploration as well.We can further enhance the performance of the saliency analysis model based on the following three points. First, minimize the number of comparisons to increase speed. Second, consider various factors comprehensively to compute saliency more scientificly. Third, introduce the appropriate post-processing steps to enhance the detection effect.This paper designs a novel clustering algorithm HAIC(Hexagonally Arranged Iterative Clustering), which introduces an appropriate pre-processing step for saliency detection, and solves the problem that global saliency analysis has good effect but its computational cost is large. HAIC algorithm clusters pixels in the CIELAB color space and flat space, and it updates the processing granularity of the input natural image to uniform superpixels with excellent adjacency relationship and fit color edge in linear time complexity.This paper introduces a new method SGC(Superpixel Global Contrast) for salient region detection. First, it does pretreatment and decomposes the given image efficiently via HAIC algorithm. Follow-up processes are all based on superpixels produced by pretreatment, so they are all very rapid. Then it focuses on the global contrast, uses a simple and efficient assessment Euclidean distance, and reduces the elements of contrast measure set to two, color difference and spatial coherence, which can not only be defined visually, but also can calculate saliency with high-precision. Finally, it introduces POO(Pivotal Object Outstanding) algorithm and processes saliency map further, to strengthen pivotal object and enhance detection effect. Compared with ten other saliency analysis models mentioned previously, SGC method generates saliency map of higher quality, and its average detection time is shorter.This paper also applies extracted saliency maps to a lot of interesting applications. Reproduce some classic application models, such as ergodic threshold extraction. Improve some representative application models, such as object optimization extraction. Design some new application models, such as object-sensitive zoom. This paper does a large number of experimental evaluations to show outstanding performance and excellent versatility and practicality of SGC.
Keywords/Search Tags:Saliency analysis, Superpixel, Global constrast, Pivotal object focusing
PDF Full Text Request
Related items