Research On Collaborative Saliency Detection Algorithm Based On Group Collaborative Learnin

Co-saliency detection is an important task in the field of computer vision.It aims to detect and segment common and attractive targets from a set of related images,so as to screen and extract useful information in the image group.The difficulty of co-saliency detection is how to model the consistency relationship between different images in the image group and effectively mine the common saliency information of the image group.This paper studies the co-saliency detection algorithm based on group collaborative learning,and explores the common saliency information of image groups through collaborative learning between two groups,the main content is as follows:（1）Aiming at the problems of common saliency target false detection and target edge in the current co-saliency detection algorithm,this paper proposes a co-saliency detection algorithm based on intra-group consistency and comparative learning（ICACL）,which uses two different image groups to form group collaborative learning.By calculating the similarity between all image semantic features in the image group,the consistency features of the image group are extracted.Two different image groups and their consistency features are used to construct positive and negative samples for comparative learning to adjust the consistency features,which reduces the false detection of common saliency targets.In addition,a feature enhancement module based on attention mechanism is designed to enhance the common saliency features,and the encoder and decoder are connected through the feature fusion module to realize the edge optimization of the common saliency target in the multi-scale feature fusion supervision process.The experiments on the three benchmark datasets show that the ICACL algorithm have advanced performance and strong robustness.（2）Based on the ICACL algorithm,the existing saliency detection algorithm is introduced to predict the saliency map of the image,and a co-saliency detection algorithm for fusing image saliency information（FISI）is proposed.The image group and its saliency map group are used to form group collaborative learning.The saliency map predicted by the saliency detection algorithm extracts the saliency information of a single image,and calculates the correlation between the semantic features and saliency information of all images in the image group to extract the consistency feature of the image group.Positive and negative samples are constructed in a group of images for comparative learning to adjust the consistency features and suppress the impact of noise objects on the output.The experiments on three datasets showed that the FISI algorithm improved their performance further on the basis of the ICACL algorithm,especially the improvement of the F_β^max metric by 1.8%on the COCA dataset.