基于U-net模型的航拍图像去绳带方法

光启云端号平台是用电缆绳索牵引气囊的, 由于空中航拍摄像系统悬挂在气囊下方, 摄取的图像不可避免地含有绳带信息, 这些绳带信息影响图像质量, 所以在场景分析和目标检测中需要剔除。提出一种基于U-net模型的绳带检测算法, 引入深度可分离卷积提高计算速度, 采用一种带权重的交叉熵作为损失函数, 解决类别不均衡带来的收敛不稳定问题, 最终的模型能够用较少的样本在较短的时间内, 快速准确地检测绳带, 利用快速行进修复算法(FMM)对绳带图像进行了修复。实验结果表明:该算法的mIOU达到62.8%, 得到了较好的去绳结果。

Aerial image de-roping based on U-net model

The Kuang-chi cloud platform is used to pull the airbag with a cable rope. Since the aerial photography system is suspended under the airbag, the captured image inevitably contains the rope information, which affects the image quality and needs to eliminated in the scene analysis and object detection. A rope detection algorithm based on U-net model was proposed, which introduced the separable convolution to improve the detection speed and adopted a weighted cross entropy as a loss function to solve the problem of convergence instability caused by category imbalance. The final model was able to quickly and accurately detect the rope with fewer data in less time, and then the rope image waseliminated using the fast marching method algorithm (FMM). The test results show that the mean intersection over union(mIOU) of the algorithm reaches 62.8% and has obtained good results.