城市场景分割的多尺度感知融合网络研究

Research of Multi-Scale Perceptual Fusion Network for Urban Scene Segmentation

针对道路场景信息多尺度变换的问题,基于编码器-解码器的非对称网络结构,提出一种轻量级多尺度感知融合网络。根据残差网络以及空洞卷积的概念,设计一种新的残差模块Res-SS,在不增加卷积参数的情况下,提高特征提取的效率。设计多尺度感知融合提取模块,提高网络对于道路场景多尺度物体信息的自适应提取能力。为弥补特征提取过程中的低级特征缺失,采用Superpixel模块,将道路场景内低级边缘信息与高级语义信息融合,使得二者互为补充,从而得到高质量的语义分割结果。在Cityscapes数据集上的实验表明,该算法比现有的轻量级城市场景语义分割算法具有更高的精度和鲁棒性。

In order to solve the problem of multi-scale transformation of road scene and adapt to the requirements of automatic driving semantic scene,and reduce the complexity of the whole structure of convolutional neural network model,this paper propos-es a multi-scale perceptual fusion semantic segmentation network based on asymmetric network structure of decoder to segment road image.According to the idea of residual network and space convolution,a new Res-SS residual module is designed to improve the efficiency of feature acquisition.The multi-scale perceptual fusion extraction module is designed and adopted to extract more multi-scale feature information from different receptive fields for weighted fusion,so as to improve the robustness of the network.Be-cause the edge information of the segmented object is lost in the process of feature extraction,a Superpixel segmentation module is used to fuse the low-level information with the high-level information,so as to recover the lost information of the feature map.Exper-iments on Cityscapes dataset show that the algorithm has higher accuracy and robustness than the existing semantic segmentation al-gorithms.