Research On Salient Object Detection Via Attention Edge Interaction

Salient object detection(SOD)aims to locate and segment the objects/regions in images that attract the most human attention.As an important image processing method,SOD is widely used in various computer vision tasks which include semantic segmentation,multimodal matching and image retrieval,etc.In recent years,the emergence of fully convolutional neural networks has greatly contributed to the development of salient object detection.To further improve the detection performance,attention and edge awareness mechanisms are employed by some state-of-the-art methods to focus on salient features and reduce the blurred boundaries of salient objects,however,the previous methods only use attention and edge awareness singly,ignoring their interrelationships.Therefore,it is still a challenge to make full use of attention and edge awareness mechanisms to improve salient features and thus detection accuracy.This paper focuses on two different types of datasets,natural images and optical remote sensing images,to research salient object detection algorithms.Specifically,the main research work and contributions of this paper are as follows:(1)Salient object detection algorithm via multiscale attention-edge interaction refinementTo enhance the interaction between attention and edge awareness mechanisms,a multiscale attention-edge interaction refinement network is proposed in this paper.The proposed model consists of two parallel and interacting subnets to achieve SOD and salient edge detection(SED),respectively,and each subnet consists of multiple interactive refinement modules cascaded in series with the multi-scale attention refinement module proposed for the SOD subnet to provide edge-enhanced attention and the edge refinement module proposed for the SED subnet to provide attentionenhanced edges.Furthermore,a progressive feature concentration structure is proposed to reduce information loss in the process of feature fusion.Compared to 20 state-ofthe-art models,the proposed model achieves highest structural similarity for all 6benchmark datasets and computationally efficient require only 26.61 M parameters and19.22 G FLOPs.(2)Salient object detection algorithm via attention-edge interaction in optical remote sensing imagesTo improve the detection performance of SOD in optical remote sensing images,this paper proposes an attention-edge interaction network,which consists of two interactive decoding branches based on U-shaped architecture to achieve SOD and SED,respectively,and a multi-scale attention interaction module is proposed to provide edgeenhanced attention and attention-enhanced edge interactively between the two branches.Moreover,to alleviate the semantic dilution problem in the feature fusion process,semantic-guided fusion module is proposed to enhance the propagation of semantic information.Through extensive quantitative and qualitative comparisons,the proposed model outperforms 15 state-of-the-art SOD methods,which further demonstrates the performance advantage of attention-edge interaction.