Multiresolution Refinement Network for Semantic Segmentation in Internet of Things

With the large-scale deployment of the Internet of Things (IoT), the demand for real-time perception and environment understanding in road scenarios is becoming increasingly urgent. Meanwhile, semantic segmentation has been widely studied as pixel-level scene parsing. However, the resource-constrained IoT equipment should consider real-time and accuracy performance during semantic segmentation. In this paper, we introduce a swift and efficient semantic segmentation network for road perception in traffic scenarios, deployed on the cloud server. For this purpose, we propose an effective and efficient semantic segmentation network, termed Multi-Resolution Refinement Network (MRRNet). It employs an encoder-decoder architecture with effective communication of branch features. In the encoder stage, a Semantic Reconstruction Module (SRM) is integrated to capture scale feature information and refine the performance of semantic features. To address the issue of information loss and bolster the representation of features, the Multi-Scale Feature Polishing Module (MSFPM) is proposed. Additionally, a Symmetric Aggregation Interaction Module (SAIM) is devised to harness the inherent complementarity between low-level and high-level features. Extensive evaluations demonstrate that MRRNet achieves competitive results on the Cityscape, CamVid, and NightCity datasets. Specifically, by employing a single RTX 3090 GPU, MRRNet-S achieves 78.7% mIoU with inference speed of 112.6 FPS on the Cityscapes dataset, 78.9% mIoU with inference speed of 175.7 FPS on the CamVid dataset, and 54.2% mIoU with inference speed of 178.9 FPS on the NightCity dataset. Our code is publicly available at https://github.com/WonderKing123/MRRNet/.