Image Saliency Detection Based On Regional Label Fusion

Image saliency detection acquires important information of image by imitating the visual attention mechanism to improve the efficiency and accuracy of image processing.It is widely used in target recognition,image annotation and retrieval,image automatic cropption,image compression and other fields.Aiming at the problems of boundary ambiguity,unclear contour and poor internal structure in the existing saliency detection methods,this paper proposes an method of image saliency detection based on region label fusion.Firstly,using the superpixel segmentation method to segment the input image,in order to reduce the number of over-segmented regions,the superpixel regions is quadratic clustered by spectral clustering method to obtain the region boundary information and color space distribution image.Then,using the conditional random field theory combine the three saliency images of the color space distribution image,the multi-scale contrast image and the center-surround histogram to obtain a coarse saliency image;In addition,a label indication vector is defined according to the region boundary information,in order to better integrate the boundary information into the coarse saliency image,the regional label fusion algorithm is designed,and the reconstructed coarse saliency image is obtained by the unified label.Finally,the reconstructed coarseness saliency image is refined by an optimized background prior method.This algorithm randomly extracts 500 images from MSRA10 K and ECSSD datasets for saliency detection,and subjective comparison and objective comparison with 11 saliency detection methods.The experimental results show that the region label fusion algorithm has a clear boundary contour,fine structure of the region and.At the same time,the method has high performance in suppressing the background and saliency detection accuracy.Moreover,the proposed method not only has a good detection effect on images with only one salient object,but also has stable accuracy and integrity on images with multi-object images and images with high internal color contrast for salient objects,indicating that the method is more applicable and robustness.This paper has 18 figures,3 tables and 60 references.