Co-Saliency Detection Based on Multi-Scale Feature Extraction and Feature Fusion

Abstract

In this paper, we propose a co-saliency detection algorithm based on multi-scale feature extraction and feature fusion. The algorithm extracts multi-scale features of images based on image information and combines these multiscale features with single image saliency maps (SISMs) generated by the edge guidance network (EGNet) to obtain single image vectors (SIVs). Based on these features, self-correlated features (SCFs) and rearranged self-correlated features (RSCFs) are calculated, and co-saliency attention (CSA) maps are created by weighting. Finally, the decoder receives the rearranged self-correlation and co-saliency maps in order to generate the final prediction maps. It can effectively solve the problem of poor performance of current feature extraction and saliency detection algorithms in complex scenes with multiple saliency targets. The simulation results show that the proposed algorithm not only improves the accuracy of co-saliency detection of RGB images in complex scenes but also reduces the error, and its performance is better than other algorithms.