Contextual Attention Refinement Network for Real-Time Semantic Segmentation
Shijie Hao, Yuan Zhou, Youming Zhang, Yanrong Guo
Abstract
Recently, significant progress has been made in pixel-level semantic segmentation using deep neural networks. However, for the current semantic segmentation methods, it is still challenging to achieve the balance between segmentation accuracy and computational cost. To address this issue, we propose the Contextual Attention Refinement Network (CARNet). In this method, we construct the Contextual Attention Refinement Module (CARModule), which learns an attention vector to guide the fusion of low-level and high-level features for obtaining higher segmentation accuracy. The CARModule is lightweight and can be directly equipped with different types of network structures. To better optimize the network, we additionally consider the semantic information, and further introduce the Semantic Context Loss (SCLoss) into the overall loss function. In the experiments, we validate the effectiveness and efficiency of our method on several public datasets for semantic segmentation. The results show that our method achieves a good balance on accuracy and computational costs.