Study On Issues Of Saliency Detection With Visual Attention And Pixel-Level Multifocus Image Information Fusion

With technology booming,the persuit of high-calibre digital media based scenario gradually reveals.Thus visual information processing has been an active research area since then.Saliency detection with visual attention and multi-source image fusion,two indispensable components of this field,carve out an significant position in all walks of life.Saliency detection with visual attention is a versatile tool applied in various aspects involving image matching,object tracking,dynamic lighting as well as video compression encoding.As a focal issue of multi-source image fusion,pixel-level multifocus image fusion plays a irreplaceable role across industries by compensating for the defect in preserving scene contextual information intact due to the field depth of imaging devices.Overall,the specific contents of this paper are demonstrated as belows:To begin with,by providing a brief overview of the characteristics and current situation about saliency detection with visual attention and multi-source image fusion,we sum up several related models so as to seek the bottleneck hindering the performance by methods of induction and analogy.Secondly,derived from the theory of mean shift clustering and of neighborhood inconsistency as well as of improved laplacian pyramid transform reconstruction,this paper proposes a novel algorithm of saliency detection with visual attention named LNCR.The fact that LNCR is capable of suppressing and weakening oscillatory components such as background noise grants robustness against complex background.In general,LNCR can keep edges and contours information nearly intact while extracting saliency region characteristics with few background noises and good visual effects.Experiments about seven well performed models(SR,SF,HC,LC,RC,FT,GC)as well as LNCR are conducted on two universally accepted public benchmark datasets of which the outcomes demonstrate the superiority of our proposed algorithm LNCR from both subjective and objective aspects.Thirdly,FASSL,a novel algorithm of pixel-level multifocus image fusion is proposed partially based upon LNCR introduced in previous chapter.In this model,the given scene images are decomposed into two components including visual layer and vein layer.As for each layer,a particular fusion rule is meticulously selected based on its morphological characteristics.The model LNCR is applied to fuse visual layers and the model SCM for vein layers.Despite the fact that the fusion of multifocus image is widespreadly applied,few researches have been conducted on a specific phenomenon that the smooth visual layer in the vicinity of edges often presents itself a distinct feature opposite to the rest of itself.To tackle with this issue,a map of affecting domain correction ADCMap is utilized to identify the peculiar region.The ameliorated algorithm applicable to both gray and color images performs well both in integrating complementary data and preserving saliency information of pixel-level multifocus images compared with seven classic image fusion models including SIDWT,DWT,FSD,NSCT,LP,GFF and IMF on four groups of horizontal comparative experiments.Comparison and evaluation of the experiment results demonstrate the superiority of our proposed algorithm FASSL from both objective and subjective aspects.Ultimately,the final chapter provides a brief overview of the research content and a general review of the innovations.Meanwhile,a prospect is forecasted of moving objects saliency detection and multifocus image fusion of dynamic scene.