Image Salient Object Detection Based On Multi-Level Feature Refinement

Salient object detection aims to simulate the human visual mechanism to detect and segment the most attractive objects in the screen.Thanks to the emergence of fully convolutional neural networks,salient object detection methods based on deep learning have made great progress.When dealing with complex or low-contrast scenes,RGB-based saliency detection methods do not predict the saliency map well.As a complementary cue to RGB images,depth maps contain rich geometric and structural information,which can intuitively describe the shape and location of salient objects and further improve the detection performance.For this reason,many RGB-D salient object detection methods have been proposed to solve the problem of poor detection results of RGB-D detection methods in complex scenes.However,in RGB-D salient object detection,the fusion and refinement of multi-level features has always been a challenge.In order to solve the problems in current research,this paper proposes the RGB-D salient object detection model based on multi-level cross-modal feature interaction and the RGB-D salient object detection model based on cross-modal and multi-level feature refinement.(1)A RGB-D salient object detection model based on multi-level cross-modal feature interaction is proposed.In order to filter the noise and redundant information of the depth map,promote the cross-modal feature interaction at different levels,and accurately highlight the salient objects in the RGB image,an effective feature interaction module is designed.Then,the fused features after feature interaction are fused into the decoder layer by layer to realize the refinement of multi-level features.Compared with 12 state-of-the-art methods in the past two years on five public datasets,the detection performance is better,which proves the effectiveness and superiority of the method.(2)A RGB-D salient object detection model based on cross-modal feature interaction and multi-level feature refinement is proposed to deal with the problems of cross-modal feature interaction and multi-level feature fusion in existing methods.Moreover,a multi-level feature fusion module is designed,which uses the skip connection to fuse the context information of multi-level features,so as to realize more effective information transmission through a progressive refinement.The experimental results of the model proposed in this chapter and 17 advanced methods on five benchmark data sets show that the method has better performance advantages.