基于粒子群算法的多尺度反卷积特征融合的道路提取

研究基于传统FCN算法下的不同比例的多尺度特征融合对于复杂场景下道路提取准确度的提高.针对复杂的航拍道路场景,设计了针对于农田环境下的FROBIT农田道路数据集,并使用全卷积神经网络(FCN)对FROBIT农田道路数据集和Massachusetts城市道路数据集进行道路提取工作.本文基于传统的FCN的网络,对其反卷积方式进行改进,采用粒子群算法(PSO)设计了不同比例的多尺度特征融合.通过将本文提出的Multi-Scale FCN网络与传统的FCN神经网络在FROBIT数据集和Massachusetts道路数据集上进行对比实验,结果表明Multi-Scale FCN网络相比于传统的FCN神经网络在提取精度上得到了提高.

Road Extraction Based on PSO Different Ratio Deconvolution Feature Fusion

To improve road extraction accuracy in complex scenes based on traditional FCN algorithm with different scales of multi-scale feature fusion,several works were carried out for the complex aerial road scene, designing a FROBIT farmland road dataset for farmland environment, extracting the road information from FROBIT dataset (farmland road) and Massachusetts road dataset (city road) based on full convolutional neural network (FCN), improving the deconvolution method based on traditional FCN network, implementing multi-scale feature fusion with different proportions based on particle swarm optimization (PSO). Comparing the multi-scale FCN network proposed in this paper with the traditional FCN neural network on the FROBIT dataset and the Massachusetts road dataset, the experimental results show that the multi-scale FCN network is superior to the traditional FCN neural network in extraction accuracy.