Research On Collaborative Saliency Detection Algorithm Based On Feature Matching Mechanis

The goal of co-saliency detection task is to detect and segment common and attractive objects or regions in a group of images.The key problem to solve this task is how to model the picture groups with different styles,sizes and noises,and effectively capture the co-salient clues.Early methods usually mine the spatial information and consistency information across the graph from bottom to top based on the low-level features made by human.Then they score each sub-region of the picture group,so as to segment the co-salient region.These methods largely rely on human prior knowledge and manual features,so they are difficult to generalize to complex scenes and still have stable performance.Recently,methods based on deep learning have shown remarkable potential.The packet synchronization detection framework developed based on u-net can automatically learn and extract high-level semantic features from images,and establish a collaborative relationship model for these features,so as to obtain very reliable results.However,there are still some problems in the co-saliency detection algorithm based on deep learning.The key problem is that the current algorithms focus on summarizing the core features of co-salient targets to find cooperative regions,so they lack the ability to capture the corresponding relationship at the pixel level.Also it is difficult to explain the meaning of each sub-region of co-salient targets.In some scenes where the difference between core features and edge features is too large,the current advanced algorithms will fail.In order to solve these problems,this paper studies the co-saliency detection algorithm of deep convolution neural network based on feature matching.The main contributions are as follows:This paper proposes a deep network architecture based on fused Gromov-Wasserstein distance and semantic guidance.Fused Gromov-Wasserstein(FGW)distance is a notation of distance among metric measure spaces and being built between different geometric domains.FGW distance can measure distances between pairs of nodes within each domain,as well as V measuring how these distances compare to those in the counterpart domain.Because GW distance can extract soft matching in the presence of diverse geometric structures and has achieved great success in discovering the correspondence with shared(semantic)structure between the source domain and target domain,we use FGW distance to capture the pairwise correspondence of each pixel feature between the target image and source image in the group.To reduce the misleading in structure matching and increase the generalization of the model,we design a semantic-aware common-attention module.The module predicts the semantic types of co-salient objects and then obtains the mask of co-salient objects through fusion semantic response by semantic type.Then,this information is modulated into the feature representation to refine the local semantic region.We verify the effectiveness of our model in the three largest and most challenging datasets,including Cosal2015,COCA,and Co SOD3 k.Our method is superior to the model published at the same time and achieves the most advanced performance.The performance of the above methods depends on the ability of the capturing collaborative features of FGW matching layer.However,the FGW distance is difficult to distinguish objects with similar structure and little semantic difference in the matching process,which limits the performance and generalization ability of the above methods.Secondly,the way of exploring co-salient clues in the above methods is very single,which also makes the semantics of collaborative features extracted from the network not comprehensive enough.Then,on the basis of the above methods,this paper tries to solve the problems existing in the advanced methods,and propose the co-saliency detection algorithm based on Cooperative aggregation and earth mover’s distance(EMD).EMD is used to measure the distance in the same distribution and match two sets of data from the same type.Therefore,we use EMD to mine cross graph collaborative clues and single saliency clues.Also,we use fourier adaptive pooling to get the embedding features of co-saliency objects.It can adaptively retain the medium and low-frequency information of the images through the frequency and corresponding amplitude,which can make the information of embedding features richer and suppress background noise meanwhile.Our method achieves better results in three datasets,especially F-measure tesing in the dataset Cosod3 k is 2.6% higher than the original model.