Research On Salient Object Detection Algorithms For Optical Images

With the increasing popularity of intelligent acquisition equipment,optical image data have shown explosive growth.The research on salient object detection algorithms for optical images aims at using computers to simulate human visual attention mechanisms,and then segment and extract the most attractive regions in optical images.This paper focuses on the scientific issue of the saliency perception mechanism in optical images.Starting from natural scenes and more complex remote sensing scenes,this research accurately obtains key information from optical images,effectively removes redundant information,and ultimately improves the ability to express saliency information in each image.Therefore,this research can efficiently use limited computing resources to process massive optical image data and can provide powerful help for advanced tasks of computer vision,such as cognitive computing,foreground extraction,semantic segmentation,change detection,and so on.The research on salient object detection algorithms for optical images mainly includes RGB salient object detection in natural scene images(RGB SOD),RGB-D salient object detection in natural scene images(RGB-D SOD),salient object detection in optical remote sensing images(RSI-SOD),and saliency region-guided change detection.The key issues are as follows:(1)The low-level feature extraction in RGB SOD;(2)The cross-modal feature utilization in RGB-D SOD;(3)The feature representation and constraint in RSI-SOD;(4)The difference feature enhancement in change detection.To address the aforementioned key issues,this paper has conducted in-depth research.The main research contents are summarized as follows:(1)Hierarchical edge refinement network for RGB SOD.To alleviate the issue of inaccurate edges and details caused by the inadequate extraction of low-level features.A coarse-to-fine detection method has been proposed,in which the saliency prediction network is used to roughly detect the saliency regions,and the edge preserving network is used to accurately detect the saliency edges.Different from the previous indiscriminate supervision strategy,a novel one-to-one hierarchical supervision strategy is adopted to supervise the different outputs of the entire network.Extensive experiments are conducted on 5 public RGB SOD datasets,and the results show that the algorithm can effectively extract low-level features,which is conducive to obtaining accurate semantic classification and clear edge detection.(2)Regional attention-guided fusion network for RGB-D SOD.To alleviate the issue that RGB and depth features are not fully utilized and the complementarity of cross-modal features cannot be effectively mined.The regional attention-guided fusion method has been proposed to fuse cross-modal features,which can maintain accurate saliency regions and abundant details.The feature fusion attention module is also introduced to emphasize the high response channel of the two-stream high-level feature fusion and filter useless features.Extensive experiments are conducted on 5 public RGB-D SOD datasets,and the results show that the algorithm can fully utilize the complementarity of cross-modal features and achieve accurate saliency prediction.(3)Adjacent complementary network for RSI-SOD.To alleviate the issue of insufficient feature representation and supervision constraints caused by the ignorance of adjacent feature fusion and global constraints.The adjacent complementary method has been proposed to complement the features among the adjacent layers and the current layer,which can effectively aggregate saliency features.In addition,the global constraints are introduced to supervise the prediction objects with various scales.Extensive experiments are conducted on 2 public RSI-SOD datasets,and the results show that the algorithm can effectively utilize the adjacent layers and add appropriate global constraints.Thus,it can achieve accurate prediction in many poor situations,such as complex backgrounds and objects with various scales.(4)Context and difference enhancement network for change detection.To alleviate the issue of inaccurate changing regions caused by insufficient spatial attention on the difference features of bitemporal images.The saliency region-guided content difference enhancement method has been proposed to enhance the change features of bitemporal images,which can make the change region more complete.In addition,the Transformer module is introduced to deeply mine the context relationship of bitemporal features.Extensive experiments are conducted on 2 public change detection datasets,and the results show that the algorithm can enhance the content difference of the change region,and fully utilize the global context relationship.Therefore,it can improve the performance of change detection.