Research On Salient Object Detection Methods Based On Deep Learning

Salient object detection is to detect the contour of the foreground from an image.It is useful with the help of deep learning methods,and is the first task of computer vision.Because more and more tasks need to rely on salient object detection,it becomes a hot research direction of artificial intelligence in recent years.The existing methods are very effective in detecting simple scenes.However,for some complex scenes,such as irregular object,many object,small object in the image,or the boundary contour of the object which is complex,some methods cannot effectively deal with these problems.The saliency maps predicted by many methods have the phenomenon that the object is bluring,the small object is losing,and the object boundary is not sharp enough.In view of the above problems,this paper focuses on the following three aspects:First of all,this paper proposes multi-scale module used to extract features to obtain rich context information,which can effectively alleviate the problem of the boundary discontinuity and the fuzzy object.Then,this paper puts forward a feature fusion module,which combines with the global context feature,the low-level feature and the high-level feature,it can not only suppress the transmission of background noise,but also recover the spatial detail structure information of the salient object more effectively.Secondly,for the problem of salient object detection in complex scenes,this paper introduces the deformable convolution to extract the feature information of irregular object objects well and sharpened the boundary of object objects in complex scenes.Then a global context attention module is proposed which can not only make up for the problem of high-level semantic information dilution but also it can enhance the important information of small objects,so that the prediction of the saliency maps in decoding stage is more accurate.Finally,in the task of salient object detection,depth information is considered to be a supplement to RGB data.In order to make full use of depth information,this paper proposes an attention mechanism modal fusion RGB-D detection method based on the RGB dataset method.This paper introduces a channel spatial attention mechanism module to filter the redundant and noise information in the low-quality depth map for matching of the two modal features.Then feature fusion RGB high-level semantic feature information and depth high-level semantic feature information can use the complementary information of the two modes better to generate global context information and enter the decoding stage.We test two types of datasets,one is RGB dataset and the other is RGB-D dataset.On the RGB dataset,this paper tests five commonly used datasets and compares them with the existing fourteen mainstream deep learning salient object detection methods.On the RGB-D dataset,this paper also tested five datasets and compared them with the existing thirteen mainstream salient detection methods.In this paper,eight evaluation indexes and the generated saliency map are used to comprehensively compare the results of this method and other methods.The experiment shows that: contrast with other mainstream detection methods,the edge contour continuity of saliency map detected by our method is better,the spatial structure details information is clearer and it is closer to the truth map.The comprehensive index,enhancement measure,average absolute error,weighted F-value,structural measure,average comprehensive index,PR curve and Fmeasure curve were significantly improved.