Research On Depth Information Estimation And Visual Saliency Based On The Mapping Relationship Between Images

Human visual system can process complex scenes in real time,so that our attention is focused on the area of interest,while ignoring the area of uninterest.In the field of computer vision,we call these areas which attract human vision as salient areas.Depth information plays an important role in distinguishing salient foreground from cluttered background in the saliency detection of RGBD images.As a supplement to color information,the quality of depth information directly affects the subsequent saliency detection results.Due to the limitation of depth information acquisition equipment and human factors,the quality of depth information obtained in different scenarios varies greatly.To solve this problem,this paper attempts to estimate a high quality additional depth map.As a supplement to the original depth information,the new depth map will be input into the newly designed selective fusion network to improve the performance of saliency detection.In order to achieve the above objectives,this paper first finds a set of images which is similar to the input image,then establishes the local and regional mapping relationship between the similar image and the input image,and then migrates the depth information of the similar image set through the corresponding relationship between the images,and obtains the roughly estimated depth map.Next,this paper constructs a fine object-level correspondence,and further improves the quality of the new depth map by using the salient prior knowledge.Finally,the original depth information and the newly estimated depth information are input into the newly designed selective depth fusion network to complete the saliency detection of depth images.The main contributions of this paper can be summarized as follows:(1)A multi-scale algorithm is proposed to calculate a reliable estimated depth map.When neither RGB information nor camera depth information can separate a salient object from its nonsalient neighborhood environment,the newly estimated depth map of this paper can make up for the lack of information in this aspect.(2)A new depth-selective salient fusion network is designed to achieve the best complementary state between RGB information,original camera depth and newly estimated depth information.