Research On Salient Object Detection Methods Based On Propagation Model

Human visual system is not only able to locate interesting regions rapidly,but also detect and recognize conspicuous objects,namely salient object,in the scene.With the development of technology and the popularity of the internet,especially the improvement of data collection techniques,the number of videos and images on the network are explosive growth.Therefore,how to extract valuable visual information from a large amount of images and videos,including object detection and recognition,is an important issue in the field of computer vision.To make computers realize this function in a rapid and intelligent way,salient object detection has gained more and more attention in recent years.Salient object detection algorithms assign limited com-putation resources to important area in the scene to provide effective information for subsequent visual tasks and improve their performance,so as to eliminate the interference of redundant areas and save computational expenses.Therefore,as an image pre-processing algorithm,salient ob-ject detection has been applied to a variety of computer vision tasks successfully,such as image retrieval,visual tracking,image reduction,etc.Considering that images contain a large amount of redundant information,the algorithm can hardly detect entire salient objects rapidly and ac-curately.However,it is an effective strategy to extract key information(e.g.,features,salient locations)in the image first and then propagate the information of key positions to other regions according to the correlation relationship between features.Therefore,research on effective and robust key information extraction method and feature propagation theory are crucial problems to be solved in this paper.The main contents of this paper are as follows:(1)Pattern mining based salient object detection algorithm.A pattern mining based salient seeds selection method is proposed.Specifically,first the pattern mining algorithm is utilized to learn the saliency patterns,which are common features of salient regions,but rarely observed in the background area.Based on saliency patterns,a large number of confident salien-cy seeds can be detected effectively,so as to remove inaccurate predictions in the initial saliency maps.Finally,an Extended Random Walk(ERW)algorithm is proposed to propagate label infor-mation of saliency seeds to other image regions to achieve global predictions of saliency values.ERW algorithm is consist of the classic random walk algorithm,a quadratic Laplacian term and an external classifier.The quadratic Laplacian term is able to extend the propagation range of the random walk,and enforces distant regions are affected by information of saliency seeds.The external classifier makes full use of known saliency information.The proposed method takes a color image and the corresponding initial saliency map generated by existing methods as input,and is able to promote the performance of existing saliency models by a large margin.(2)Exemplar-aided salient object detection via joint latent space embedding.First a confident saliency seeds detection method is proposed,then saliency information is propagated to the global area.Specifically,we utilize annotated reference images to assist the detection of salient objects.First,a dictionary learning based latent space embedding algorithm is proposed to map both the reference image and the test image into a joint latent space,and then the label transfer process is conducted,so as to obtain a large number of confident saliency seeds of the test image.Subsequently,the random walk algorithm is employed to propagation information of saliency seeds to other regions.In this work,performance of unsupervised saliency models can be improved by using very few annotated reference images.This not only solves the problems that supervised algorithms need large amount of annotated training data set and expensive train-ing time,but also improves the issue that unsupervised algorithms rely on heuristic assumptions of salient objects which introduce noises and errors.(3)Spatial context-aware network for salient object detection.A feature propagation algorithm on the deep feature level is proposed to explore global and local contextual informa-tion of the input image.Specifically,a spatial context-aware network is constructed,in which global contextual information is explored by a long-path context module which is based on the global propagation model and endows discrimination ability to coarse scale feature maps;lo-cal contextual information is explored by a short-path context module which is based on a local propagation model.The short-path context module integrates feature maps from multiple scales seamlessly,and compensate detailed saliency cues to prediction results progressively by discov-ering local contextual information.The proposed feature propagation based multi-scale fusion strategy has the advantages of enhancing the capability of feature representation and promoting the interaction between global and local information.What is more,it avoids noises and cluttered information introduced by empirically feature integration methods.Experimental results on five large-scale benchmarks show that the proposed method achieves the favorable performance and outperforms the state-of-the-art algorithms consistently.