Research On Image Co-Salient Object Detection Algorithm Based On Deep Learning

Image co-salient object detection task is a branch of image salient object detection,and it is a research hotspot in the field of computer vision.The task of image co-salient object detection is to detect common and attractive objects or regions from a group of related images.The key to solving this problem is how to extract the consensus clues of co-salient objects and avoid the interference of irrelevant salient objects.Early traditional methods usually use handcrafted low-level features to score similar regions of images,and then segment common regions.Traditional methods mostly rely on people’s priors,which are difficult to apply to real-world scenarios.Recently,the rapid development of deep learning has promoted the co-salient object detection technology under deep learning and exhibited impressive potential.Existing methods for image co-salient object detection based on deep learning often have the following shortcomings:(1)group semantic features are not fully explored;(2)contextual information of common objects is not effectively explored;(3)neglect of co-salient objects The relationship between local features and global dependencies;(4)The consistency of the consensus within the group is not effectively maintained during the decoding process.To solve the above problems,this paper designs a two-image collaborative salient object detection method.Aiming at questions(1)and(2),this paper proposes a group semantic-guided neighbor interaction network for co-salient object detection.Specifically,the proposed network includes a group semantic module,a neighbor interaction module and a feature enhancement module.The network first learns semantic consensus from a group of related images through group semantics,and mines potential high-level semantics from forward and reverse features through a reverse guidance strategy and a channel grouping strategy;afterward,common cues are mined through global attention to obtain group Semantic consensus.Guided by the group semantic consensus,the neighbor interaction module interacts with adjacent layer features to explore contextual clues to collaborative feature representations.Finally,the feature enhancement module refines key cues through an attention mechanism,which improves the consistency and compactness of synergistic features.The proposed network achieves state-of-the-art performance on three challenging co-salient object detection benchmark datasets.In addition,extensive ablation experiments can also demonstrate the effectiveness of the proposed three modules.Aiming at questions(3)and(4),this paper proposes a co-salient object detection network based on the parallel interaction of Transformer and Convolutional Neural Network(CNN)networks.The network can efficiently mine local information and global representation for co-saliency learning through the parallel interaction of Transformer and CNN.This paper proposes three key components,namely a mutual consensus module,aconsensus complementary module,and a group-consensus progressive decoder.The mutual consensus module aims to capture the global consensus from the high-level features of the CNN branch and the Transformer branch,as a guide for subsequent integration of two-branch consensus cues at each level.The Consensus Complementary Module aims to effectively fuse local information and global contextual information from different levels of the two branches.The group-consistency progressive decoder aims to maintain the consistency of group features and predict accurate co-saliency results.Furthermore,extensive experiments demonstrate the effectiveness of the proposed method as well as the designed components.