Shadow Detection with Attention Feature Block and Multi-scale Weight Segmentation Network

Abstract

Shadow detection is an important and challenging task in computer vision. However, the current shadow detection method has limited semantic information extraction and has low detection accuracy for small shadow regions. Therefore, we propose a new shadow detection network that extracts both global semantic information and the context of the middle layer in a targeted manner. We use the single network structure of the encoder-decoder. The encoder module uses resnet-101 to extract features. The decoder module extracts semantic information by using the feature block with attention, which fuses the attention mechanism and divides the feature map, and uses the feature block after up-sampling to increase the resolution for feature extraction, so as to make it pay attention to the features of small shadow regions and improve the detection of small shadow regions. Furthermore, weight generator module and weight fusion module are used to combine the generated weight maps with the initial output of the network to further extract global context information. Finally, we use the pointrend module to complete the boundary refinement. We evaluate our network on the two benchmark datasets of SBU and UCF. The experimental results show that our method is superior to the state-of-the-art methods and the balance error rates on the two datasets are reduced by 20.0% and 10.7%, respectively.