Attention Res-Unet: an Efficient Shadow Detection Algorithm

Yue DONG,Hua-jun FENG,Zhi-hai XU,Yue-ting CHEN,Qi LI

DOI: https://doi.org/10.3785/j.issn.1008-973x.2019.02.021

2019-01-01

Abstract:Shadow pixels in images can lead to the uncertainty of image content, which is harmful to computer vision tasks. Therefore, shadow detection is often used as a preprocessing step of computer vision algorithm. A shadow detection network was proposed by combining semantic information contained in input images and correlation between pixels. Pre-trained deep network ResNeXt101 was used as feature extraction front-end module to extract semantic information of the image. The baseline structure of the network was built to up-sample feature layers, encouraged by the design idea of U-Net. Non-local operations were added before the output layer to provide global information for each pixel and establish the relationship between pixels. At the same time, an attention generation module and an attention fusion module were developed to further improve shadow detection accuracy. Two common shadow detection datasets named SBU and UCF were utilized for verification. Experiment results showed that the proposed network outperformed previous methods in both visual effect and objective indicator. The proposed network showed 14.4% reduction on SBU and 14.9% reduction on UCF for the balance error rate, compared with the state-of-the-art framework.