基于DenseNet的复杂交通场景语义分割方法

Semantic Segmentation Method for Complex Traffic Scene Based on DenseNet

针对交通场景语义分割方法存在参数量较大、计算效率较低、精度不足等问题,文中提出基于全卷积化DenseNet的多尺度端到端语义分割模型.首先,构建一种含混合空洞卷积的密集连接模块,同时沿通道维度级联各模块,用于提取图像特征.然后,采集多尺度视觉信息并以此作为监督信号回传至原通道中.最后,通过双线性插值法获得预测输出.在CityScapes数据集上的测试实验表明,文中方法对复杂交通场景的解析能力较强,预测精度和分割效率较高.

An end-to-end multi-scale semantic segmentation model based on fully convolutional DenseNet is proposed,aiming at the problems of traditional semantic segmentation methods for street scene,such as the large number of parameters and low computational efficiency and precision.Firstly,convolution layers embedded with hybrid dilation convolution are stacked to establish a dense module,and then the modules are cascaded along channel dimension to extract features.Next,multi-scale visual information regarded as supervised signals are transferred back to original channels.Finally,the prediction results are obtained by bilinear interpolation method.Experimental results on Cityscapes dataset demonstrate that the proposed method achieves an efficient segmentation and performs a better accuracy for street scene parsing.