Lightweight and efficient asymmetric network design for real-time semantic segmentation

Abstract

With the increasing demand for application scenarios such as autonomous driving and drone aerial photography, it has become a challenging problem that how to achieve the best trade-off between segmentation accuracy and inference speed while reducing the number of parameters. In this paper, a lightweight and efficient asymmetric network (LEANet) for real-time semantic segmentation is proposed to address this problem. Specifically, LEANet adopts an asymmetric encoder-decoder architecture. In the encoder, a depth-wise asymmetric bottleneck module with separation and shuffling operations (SS-DAB module) is proposed to jointly extract local and context information. In the decoder, a pyramid pooling module based on channel-wise attention (CA-PP module) is proposed to aggregate multi-scale context information and guide feature selection. Without any pre-training and post-processing, LEANet respectively achieves the accuracy of 71.9% and 67.5% mean Intersection over Union (mIoU) with the speed of 77.3 and 98.6 Frames Per Second (FPS) on the Cityscapes and CamVid test sets. These experimental results show that LEANet achieves an optimal trade-off between segmentation accuracy and inference speed with only 0.74 million parameters.