Research On High-performance Collaborative Saliency Detection Algorithm Based On Deep Learnin

The co-salient object detection task aims to detect and segment common,semantically similar salient objects or regions in a set of images.This task has a wide range of potential applications,including video saliency detection,image retrieval,collaborative localization,remote sensing images change detection and more.In the past fifteen years,co-salient object detection has made significant progress,but existing methods still have considerable room for improvement when faced with challenges such as out-of-distribution data,real-world scenes,multimodal information clutter,disguised collaborative targets,drastic scale changes,noisy background,and salient object adversarial interference.To design more robust and practical co-salient object detection models,this paper improves model performance from three aspects: unknown class generalization,image matching and multimodal alignment,and modeling uncertainty in real-world scenes.The main innovations and contributions of this paper include the following three aspects:(1)Regarding unknown class generalization,this paper addresses the differences between image foreground and background,which were previously ignored by existing methods,and the problem of overconfidence caused by using category labels as supervised signals.This paper investigates how to improve supervised learning by unsupervised clustering or using self-labeling in a self-supervised manner to alleviate overfitting to the training data.A high-order spatial semantic network modulation subnetwork and a multi-view self-labeling strategy are proposed.This paper explores the feasibility of using clustering and graph aggregation to classify image foreground and background,and studies self-labeling strategies to improve model robustness and generalization to unknown classes in a selfsupervised manner.In addition,this paper provides a detailed comparison between cosegmentation and co-salient object detection in similar and different fields.(2)Regarding image matching and multimodal alignment,this paper addresses the problem of complex nonlinear relationships intra-and inter-images that are difficult to model.The paper uses Brownian distance covariance matching to measure similarity between images and models complex nonlinear information.Also,the development of sensor devices has brought additional valuable information to many computer vision tasks,but there is a problem of misalignment between multimodal information in images.To better align multimodal information,this paper calibrates the original depth information based on physical depth features,aligns it with RGB features to highlight the correct salient object areas in the depth map,and suppresses interference areas to improve model performance.(3)This paper also notes that the group consensus assumption,which has been widely used in previous work in this field,limits the robustness of the model when facing difficult samples,especially when there are unrelated images in the input image group.This greatly limits the model’s application in real-life scenarios because in most cases,the image groups collected in real application scenarios cannot meet the group consensus assumption.In terms of modeling uncertainty in real-world scenarios,this paper designs a group exchange mask strategy and an uncertainty modeling and fusion branch,which can effectively improve the robustness and practicality of the model.