深度学习通过训练样本进行特征识别,已经被广泛应用于道路提取领域。该方法不局限于特定类型的影像,但是受训练样本数量和计算机硬件的限制,所提取的道路会有断裂和噪声。针对上述问题,使用VGG卷积神经网络对道路进行初步提取后引入张量投票方法进行优化处理。首先通过影像变换、随机裁剪、过采样等方法对样本进行多模式扩充,进而训练VGG卷积神经网络模型;其次利用该网络从原始影像中初步分割道路面,接着对道路面的二值影像进行张量投票获取道路的显著性信息;最后在特征提取时针对显著性信息加入阈值获取道路面。实验结果表明,所提方法提取道路的召回率与正确率均达90%以上,与其他传统方法相比具有更高的精度,验证了所提方法的有效性。

Deep learning is widely used in road extraction by training samples for feature recognition. This method is not limited to a specific type of image; however, the extracted road will be broken and noisy owing to the restriction of the number of training samples and computer hardware. Owing to the above problems, in this study, we use a VGG convolutional neural network to preliminarily extract roads and introduce a tensor voting method for optimization. First, multi-mode expansion of the samples is performed via image transformation, random cropping, and oversampling, and then, a VGG convolutional neural network model is trained. Second, the network is used to segment the road from the original image. Then, the tensor voting for the binary images of the road surface is used to obtain saliency information about the road. Finally, the threshold of significant information is set to obtain the road surface in the feature extraction process. The experimental results show that the recall rate and precision of the extracted road obtained by the proposed method are more than 90%, and the proposed method has a higher accuracy than other traditional methods, which verifies the effectiveness of the proposed method.