Research On Image Salient Object Detection Models And Methods

Visual saliency is basically a process that detects scene regions clearly different from their surroundings.Generic object level saliency detection is important for many vision tasks and has received increased interest from the computer vision community in recent years.Salient object detection methods can be categorized as bottom-up(stimuli-driven)and top-down(task-driven)approaches.Bottom-up results can be biased toward targets of interest by top-down cues such as object features,priors,scene context and task demands.Bottom-up and top-down factors should be combined into a unified framework for the general visual attention analysis.Although various computational models have been developed,the problem remains challenging.This thesis deeply researches on image salient object detection models and methods.The primary works are as follows:(1)A novel automatic salient object detection algorithm,which integrates context-based saliency with location computation based on the boundary priors,is proposed.Input image is expressed as a close-loop graph with superpixels as nodes and salient object of image has a well-defined graph-based manifold ranking(MR)location.The saliency of the image elements is defined based on their relevances to the given seeds or queries.Saliency object location is carried out in a two-stage scheme to extract background regions and foreground salient objects efficiently.A location weight is introduced to measure the relationship between each superpixel and the centroid of the detected salient regions to eliminate the background.Saliency map is computed through context analysis and location computing based on superpixels.(2)A salient object detection algorithm based on improved MR and bootstrap learning method is proposed.Firstly,a comparative analysis of recently proposed four saliency detection methods based on boundary priors is given.To increase the precision while preserving the high recall in salient object detection,three schemes are proposed.Firstly,simple visual features including color,orientation,and spatial information are used to represent image superpixels.Secondly,to overcome the shortage that the prior background assumption used in existing salient object detection method based on graph-based MR,the affinity propagation clustering(APC)is utilized to aggregate the superpixels(nodes)to different feature clusters adaptively.According to the modified boundary connectivity,the superpixels along the image boundaries are assigned with different background weights(the values of background probability).The real background seeds are selected and the improved MR method is employed to compute saliency.Thirdly,to further improve the performance,the result of the second step acts as the weak saliency map.The bootstrap learning algorithm is used to generate the strong saliency map and the final result.(3)A general computational framework is presented for detecting salient objects(e.g.cars and pedestrians)in images based on a novel attentional control mechanism.Its three parts(selection history,current goal and physical salience)are integrated to compose an integrative framework.The qualitative and quantitative effects of current selection goals and selection history are verified in experiments.The experimental results also indicate that there is competition among selection history,current goal and physical salience to detect correct salient objects.(4)A novel top-down fine-grained saliency detection method based on deep-learned features is firstly proposed in the fifth chapter,which can detect the same-object in input image as the query.The query image and its three subsample images are used as top-down cues to guide saliency detection.Because faces are intrinsically salient,always attract attention and gaze,the model is evaluated via face detection.This chapter also presents a multiscale deep features based visual saliency algorithm and regularized regression for refinement.In order to obtain more fine-grained saliency detection results,a saliency refinement method is presented based on graph Laplacian regularized nonlinear regression with a closed-form solution.The experimental results indicate that deep learning based saliency detection methods can substantially improve the performance.