A Lightweight and Dynamic Convolutional Network for Real-time Semantic Segmentation

Semantic segmentation is a difficult task that satisfies most of the demands of autonomous driving and drone aerial photography in a unified fashion. Convolutional neural networks can properly categorize image pixels via end-to-end model training. However, achieving the optimal trade-off between segmentation precision and the number of network parameters while maintaining a suitable inference time has become a challenging task. In this paper, we propose a lightweight dynamic convolutional semantic segmentation network, LDCNet, which belongs to the asymmetric network architecture. First, we designed a coding module that includes dynamic convolution: DDAB. The success of this module is attributed to the use of dynamic convolution, which increases the utilization of local and contextual information of the features. We also designed the decoding module containing feature pyramids and hybrid attention: HA-FP,which performs a multi-scale fusion of features accompanied by feature selection. On the Cityscapes and Camvid datasets, LDCNet obtains 73.5 mIoU and 69.4 mIoU accuracy with 78.4 FPS and 91.3 FPS, respectively, without pre-training or post-processing. Our experimental findings reveal that LDCNet achieves an outstanding balance between segmentation accuracy and network parameters with just 0.96 M.