SEDNet: Real-Time Semantic Segmentation Algorithm Based on STDC

Abstract

Recently, deep convolutional neural networks (DCNN) have been widely used in semantic segmentation tasks and have achieved high segmentation accuracy. However, most algorithms based on DCNN have high computational complexity, making them unsuitable for real-time segmentation. To solve this problem, this paper proposes a real-time semantic segmentation algorithm based on the STDC network. The algorithm adopts an “encoder–decoder” embedded in a U-shaped architecture to realize real-time segmentation while maintaining high accuracy. Following the encoder, a mixed pooling attention module is designed to expand the receptive field, enhancing the network model’s learning ability in complex scenarios. Then, a feature fusion module is used for combining features from different stages, and channel attention based on atrous convolution is employed to expand the receptive field and avoid dimensionality reduction learning. Finally, a Tversky-based detail loss function is used to encode more spatial details. The proposed algorithm was extensively tested on the challenging Cityscapes and CamVid datasets, and the experimental results showed that the proposed algorithm obtained 76.4% and 72.8% of mIoU, respectively. Meanwhile, our algorithm achieves 105.2 FPS and 165.6 FPS inference speed with a single NVIDIA GTX 1080Ti GPU, meeting the real-time segmentation requirements. The proposed algorithm can conduct real-time segmentation while maintaining high accuracy, achieving a good balance between accuracy and speed.