一种深度学习分割遥感影像道路的损失函数

传统的根据光谱特征或形态学算法来分割道路，存在精度低、阈值难确定等缺点，而深度学习中已有的方法并未考虑道路的特性，只是利用通用方法来分割道路。针对上述不足，提出了一种针对道路特有形态的深度学习损失函数——形态损失函数。首先使用连通性算法将预测结果划分为若干个相互分离的连通区域，分别计算这些区域的面积与外接圆面积的比值，然后取平均值作为此批训练数据的形态损失函数，最后将形态损失函数按一定的比例与交叉熵损失函数求和，得到最终的损失函数。通过在公开的遥感数据集上使用深度学习网络进行对比试验，附加了形态损失函数后平均交并比（MIoU）、准确度（ACC）及F1 Score均有提高。从预测的图形来看，附加了形态损失函数后，预测的道路更为连续。所提出的形态损失函数可用于提高光学遥感影像道路分割的精度。

A loss function of road segmentation in remote sensing image by deep learning

Traditional road segmentation based on spectral features or morphological algorithms has some disadvantages such as low precision and difficulty in determining the threshold value, and the existing methods in deep learning do not consider the characteristics of roads, only using general methods to segment roads. A deep learning loss function named morphological loss function with road unique trait was proposed. Firstly, the connectivity algorithm was used to divide the prediction results into several separated connected regions, and the ratios of the region area to the circumscribed circle area was calculated respectively. Then, the average value of regions was taken as the morphological loss function of this batch of training data. Finally, the morphological loss function was summed with the cross entropy loss function according to a certain proportion to obtain the final loss function. Through the comparative experiment on open remote sensing dataset, MIoU, ACC and F1-Score were all improved by the addition of morphological loss function. According to the prediction image, the predicted road was more continuous when morphological loss function was added. The morphological loss function proposed is an effective method to improve the accuracy of road segmentation in remote sensing.