LAANet: lightweight attention-guided asymmetric network for real-time semantic segmentation

With the increasing demand for real-world scenarios such as robot navigation and autonomous driving, how to achieve a good trade-off between segmentation accuracy, inference speed and model size has become a core issue for real-time semantic segmentation applications. In this paper, we propose a lightweight attention-guided asymmetric network (LAANet), which adopts an asymmetric encoder–decoder architecture. In the encoder, we propose an efficient asymmetric bottleneck (EAB) module to jointly extract local and context information. In the decoder, we propose an attention-guided dilated pyramid pooling (ADPP) module and an attention-guided feature fusion upsampling (AFFU) module, which are used to aggregate multi-scale context information and fuse features from different layers, respectively. LAANet has only 0.67M parameters, while achieving the accuracy of 73.6% and 67.9\(\%\) mean Intersection over Union (mIoU) at 95.8 and 112.5 Frames Per Second (FPS) on the Cityscapes and CamVid datasets, respectively. The experimental results show that LAANet achieves an optimal trade-off between segmentation accuracy, inference speed, and model size.