Global Multi-scale Fog Salient Object Detection Based on Local Attention Feature

Abstract

In foggy conditions, the target in the image is affected by fuzzy edge and low resolution, which limits the accuracy of target detection. In order to improve the detection accuracy of salient targets in foggy conditions, a detection algorithm with two-path global multi-scale and local attention features was proposed. Firstly, the image is defogged. Then the network model is divided into two branches after passing through the backbone network to extract the local and global features of the image respectively. At the same time, attention mechanism is introduced to enhance local features. Finally, local and global features at different scales are integrated to obtain more expressive output features. The proposed algorithm is tested on three data sets of PASCAL-S, DUT-OMRON and SOD. It was found that the maximum values of F-measure were 0.901, 0.832 and 0.896, and MAE were 0.061, 0.043 and 0.073, respectively. The results show that the proposed algorithm performs well in foggy weather compared with existing algorithms, which verifies the effectiveness of the proposed algorithm.