| With the increasing demand for image data,image processing technology is becoming more and more mature.Although the relevant technology in the field of computer vision has made great progress,the salient object detection of natural scene images is still an open and challenging problem in the real world.In many existing models,problems such as incomplete detection target and fuzzy boundary often occur in the task of salient object detection.The essence is that the network lacks of the accurate understanding of salient object semantics and the ability to refine the edge of the object.The former is directly related to the inherent defect of convolutional structure that is lack of long-range modeling ability,while the latter can be effectively improved with a larger number of parameters.In order to solve the above problems,we propose a salient object detection network based on self-attention with windows(SAWNet).The SAWNet method in this paper is a U-shaped structure based on switching convolution module,which combines U-Net and Transformer efficiently to obtain high quality network performance for saliency object detection.In order to make the network retain the local and fine modeling while introducing the global modeling capability,we adopts the swin convolutional module as the basic component of the network,including the continuous swin Transformer module which is good at non-local modeling and the residual convolutional module which is good at local modeling.The residual convolutional block should retain the features as much as possible and pay attention to the local information.As the basic unit of feature representation and remote semantic information interactive learning,continuous swin Transformer blocks can focus on non-local network information and enable the network to obtain global and local feature representation.Residual convolution module preserves features as much as possible and focuses on local information.In addition,in order to obtain smoother and complete prediction results,Pixel Shuffle was adopted in this paper for up-sampling.In this paper,a practical high-order degradation model is introduced to simulate the real image degradation process.And the input image is pre-processed.In addition,the proposed method is tested on five public data sets using five commonly used evaluation indicators,and compared with thirteen existing advanced methods to verify the effectiveness of the network.A large number of experiments show that the proposed method has obvious advantages. |
