Image Saliency Detection And Application Based On Dynamic Background Modeling

Visual saliency refers to the ability of an image area to attract visual attention because of its particularity and the difference from the surrounding areas or the subjective consciousness of the observer.Saliency detection aims to detect such kind of regions of interest in an image or video.This thesis summarizes the state-of-the-art of image saliency detection,studies saliency detection algorithm based on sparse representation theory and background dictionary learning,and combines the research results with engineering application.We summarize the main work and contributions in this thesis as the following.(1)Image saliency detection based on dynamic background modeling(SOD-DBM).First,an input image is partitioned into superpixels via multi-scale SLIC segmentation,and describes each superpixel via color and position feature extraction.By means of two-stage dynamic background modeling and "elbow" threshold filtering rule,multi-scale background dictionaries are learned.With each single-scale background dictionary,the reconstruction errors for all superpixels are estimated and smoothed,and the saliency map for each scale is created.In the end,the final saliency map is generated with the fusion of multi-scale saliency maps.With the combination of dynamic background modeling and single-scale background super-pixel recognition mechanism based on "elbow" rule,the problem of salient target detection failure caused by the "central prior" hypothesis about salient regions can be effectively avoided.Even if salient regions appear at image borders,saliency detection is still effective.The comparison experiments between the proposed SOD-DBM algorithm and 11 typical saliency detection algorithms on several image databases show that SOD-DBM algorithm can improve the recall rate of saliency detection without losing the precision rate.(2)Image retrieval based on query image salient region detection(CBIR-SOD).By combining the proposed saliency detection algorithm with the engineering application of image retrieval,we proposed an image retrieval algorithm,CBIR based on query image salient object detection(CBIR-SOD)?For a given query image,the salient object in the image is first detected and extracted for further feature description.Based on the content description of the extracted salient area,one can get the retrieved relevant images from image database.Image retrieval experiments on several categories of images show that the performance of CBIR-SOD algorithm is obviously better than then that of basic CBIR algorithms.