一种面向城市场景的轻量级实时语义分割网络

A lightweight real-time semantic segmentation network for urban scene

在自动驾驶系统、无人机、机器人和视频监控等移动终端设备中,语义分割深度卷积神经网络的复杂程度也随着网络层数的增加而增加。虽然这能更好地提高网络的精度,但是在日常生活真实场景下,需要考虑网络的参数和运算速度,并同时确保较好的网络精度。基于上述需求,开发一种对移动设备内存和计算能力固定且相对较小的实时语义分割场景理解系统是有必要的。根据先前的实时语义分割模型的启发思想,设计了一种轻量化的基于注意力机制的实时语义分割模型ALRNet(Attention-Light-Realistic Net)。通过在城市场景的2个数据集Cityscapes和Camvid上进行实验并与其他模型进行比较,结果表明,所设计的网络模型能够在提升分割速度的同时,也同样保证了分割的精度。做到了分割精度与速度的平衡。

In mobile terminal devices such as autonomous driving systems,drones,robots and video surveillance,the complexity of semantic segmentation deep convolutional neural networks also increases with the increase of network layers.Although this can better improve the accuracy of the network,in the real scene of daily life,the parameters and operation speed of the network should be considered first,and at the same time to ensure better network accuracy.Based on the above requirements,it is necessary to develop a real-time semantic segmentation scenario understanding system with fixed and relatively small memory and computing power of mobile devices.Based on the previous real-time semantic segmentation model,a lightweight real-time semantic segmentation model based on attention mechanism ALRNet is designed.The results show that the effectiveness of the network model of this design is verified by experimenting on two datasets Cityscapes and Camvid and comparing them with other existing models.The designed network model can improve the speed of segmentation while also ensuring the accuracy of the segmentation.A balance between segmentation accuracy and the speed is achieved.