Lightweight Real-Time Image Semantic Segmentation Network Based on Multi-Resolution Hybrid Attention Mechanism

Abstract

Effective perception of the surrounding environment and the balance between accuracy and processing speed are crucial for the successful application of real-time semantic segmentation algorithm in the fields of autonomous driving, drones, and smart security. In this paper, a lightweight feature reuse network MHANet for real-time semantic segmentation is proposed. The main novelties of our method are improved ResNet and attention-based fusion mechanism. And the effectiveness of our method is verified by a large number of experiments. Without any pre-training process, the performance of real-time segmentation is improved by using deep fusion of segmentation maps with different resolutions. At the same time, our network converges faster than other networks using pre-training when trained from scratch. Compared with existing methods, the results obtained with our method on the Camvid dataset improve in accuracy (mIoU) ranging from 2% to 6% and in efficiency (FPS) ranging from 15% to 18%. The results achieved 71.87% mIoU of accuracy in the Cityscapes test set, processing images at 203 FPS. Experiments show that manual designed MHANet is effective in improving the performance of real-time semantic segmentation without any pre-training.