边界感知的实时语义分割网络

目前多数实时语义分割网络不仅同时处理边界和纹理等细节信息而且还忽略了语义边界区域特征，从而导致物体边界分割质量下降。针对该问题，提出一种边界感知的实时语义分割网络，主要从三个方面提高边界语义分割质量。提出了边界感知学习机制利用位置信息降低边界特征和轮廓附近细节的耦合度使边界感知和位置关系相互促进。设计轻量级区域自适应模块增强卷积网络对复杂语义边界区域的建模能力。根据采样区域像素贡献值不同设计了高效的空洞空间金字塔池化模块以增强重要的细节和语义特征。实验方面，与基准相比，在Cityscapes验证集上精度提升了约5.8个百分点，在Cityscapes测试集上以47.2 FPS的推理速度使精度达到了74.9%。在CamVid数据集上与BiSeNetV2算法相比mIoU提升了约3.96个百分点。

Boundary-Aware Real-Time Semantic Segmentation Network

Currently most real-time semantic segmentation networks not only process detailed information such as boundaries and textures, but also ignore the features of semantic boundary regions which leads to the deterioration of the quality of object boundary segmentation. To solve this problem, this paper proposes a boundary-aware segmentation network, which mainly improves the quality of boundary semantic segmentation from three aspects. Firstly, a boundary-aware learning mechanism is proposed to use location information to reduce the coupling degree of boundary features and details near the contours, so that boundary perception and location relationships can mutually reinforce each other. Secondly, a lightweight region adaptive module is designed to enhance the ability of convolution network to model complex semantic boundary regions. Finally, according to the different pixel contribution values of the sampling area, an efficient distinctive atrous spatial pyramid pooling module is designed to enhance important details and semantic features. In experiment, compared with the benchmark, the accuracy on the Cityscapes validation set is improved by about 5.8?percentage points, and the accuracy on the Cityscapes test set is 74.9% with an inference speed of 47.2 FPS. Compared with BiSeNetV2 algorithm on the CamVid data set, the mIoU is improved by about 3.96 percentage points.