| The field of salient object detection has attracted the interest of many researchers in recent years,and many salient object recognition algorithms have been proposed one after another and have been successfully applied to the fields of video surveillance,image compression,and intelligent transportation.This paper is oriented to the application requirements of salient object detection in computer vision,we address the problem of poor performance in salient object detection caused by inadequate control of feature fusion methods in existing methods,and insufficient utilization of global and local features.Apply deep learning methods to make new attempts in feature fusion and backbone network optimization schemes and carry out research on salient object detection based on feature fusion and backbone network optimization schemes,focusing on salient object detection feature fusion methods and backbone network feature optimization schemes.The specific research content is as follows:Firstly,this paper proposed a self-refine,fusion,feedback network(SRFFNet)for salient object detection.This network mainly includes the following parts: self-refine module,feature fusion module,global optimization module,and feedback module.In particular,the self-refine module is mainly used to integrate and optimize the feature information obtained from the backbone network;The global optimization module extracts global feature information for subsequent feature fusion;The feature fusion module adaptively selects feature information for progressive fusion;The feedback module can correct the fuzzy boundary of the saliency map;In addition,this paper also proposed a weighted loss function to optimize training losses for better performance.SRFFNet can accurately segment significant target regions and provide clear detailed information.Without any fancy processing,this method can reach an advanced level in six evaluation indicators on five benchmark datasets compared to other world advanced methods.Secondly,for the backbone network feature optimization scheme,this paper proposed a fusion model Swin SOD for RGB salient object detection.It used Swin Transformer as an encoder to extract hierarchical features,was driven by attention mechanisms to bridge the gap between different hierarchical features,was guided by global information to detect salient regions,and used feedback information to refine the boundaries of salient objects.Specifically,Swin Transformer acted as an encoder to extract multi-level features;Channel recalibration module to optimize intra layer channel characteristics;The feature fusion module implements inter layer feature fusion guided by global information;In the second stage,feature fusion was guided by feedback information to achieve edge thinning.The Swin SOD model outperforms the State-Of-The-Art(SOTA)model on the existing five popular SOD datasets,indicating that the network has advanced performance.Based on the above research and driven by the current problems in salient object detection,this paper proposed two salient object detection algorithms from the level of feature fusion and backbone network feature optimization,which were of great significance for high-precision and rapid deployment of salient object detection in computer vision task preprocessing. |
