DDC-Net: Semantic Segmentation for Urban Roads Based on Improved Capsule Networks

In recent years, with the gradual progress of automatic driving technology, semantic segmentation of road scenes, as the core of this technology, has become a hot spot of research. However, nowadays, most of the convolutional (CNN)-based methods appear to be inefficient and costly due to the factors of large amount of detection data and complex structure. It limits their performance in dealing with some fast response (real-time) tasks. Addressing the above problems, this paper proposes a capsule network-based semantic segmentation method for road images, which achieves a good balance between recognition efficiency and detection speed. Specifically, the DDC-Net designed based on capsule network is used as the baseline network, and different connection paths are dynamically selected according to pixel affinity during forward propagation. In addition, DDC-S and DDC-G are designed for spatial detail fusion and semantic fusion, respectively, and the local feature extraction module (LFCE) is designed using a two-branch structure. Numerous experiments show that the method described in this paper outperforms most of the current CNN-based methods in terms of model size, recognition flexibility and overall performance. In ADE20K and Cityscapes test datasets, the method described in this paper achieves 74.5% and 79.4% mean intersection and merger ratio (mIoU) accuracies at 63.9fps and 64.8fps, and the experimental results demonstrate the effectiveness of our method.