在合成孔径雷达(SAR)图像敏感目标检测与识别领域中,机场跑道是重要的军事民用目标,在SAR飞机目标检测和机场关键信息提取等任务上具有重要意义。传统机场道路掩膜提取算法容易导致分割性能差、鲁棒性低,往往需要大量精心设计的鉴别后处理模块。针对图像分割任务,以卷积神经网络为基础的语义分割模型具备较好的实用价值,广泛用于光学自然图像分割领域,但是针对大区域SAR机场跑道提取任务,经典网络模型难以在保持图像特征感受野和保持图像细节上取得较好权衡。文中以DeepLabV3网络作为编码器,设计了基于上采样卷积模块和多低维特征融合的解码器用于输出预测概率图,然后添加了基于条件随机场的优化模块,进一步提升了分割结果的准确性。基于高分三号的实验表明,所提方法在分割性能指标上取得了最优的结果,并且能够有效保留机场边缘和图像细节。

In the field of synthetic aperture radar (SAR) image sensitive target detection and recognition, the airport road is an important military and civilian target, which is of great significance in tasks such as SAR aircraft target detection and airport key information extraction. Traditional airport road mask extraction algorithms tend to cause poor performance, low robustness, and often require a large number of fine identification modules. For image segmentation tasks, semantic segmentation models based on convolutional neural networks have good practical value and are widely used in the field of optical natural image segmentation. However, for large-area SAR airport runway extraction tasks, classic network models are difficult to maintain a better trade-off between keeping the receptive field size and keeping the image details. In this paper, the DeepLabV3 network is used as the encoder, and a decoder based on the upsampling convolution module and multi-low-dimensional feature fusion is designed to output the prediction probability map, and then an optimization module based on conditional random fields is added to further improve the accuracy of the segmentation results. The experiment based on Gaofen3 SAR data show that the proposed method achieves the best results in segmentation performance indicators, and can effectively preserve airport edges and image details.

国家自然科学基金资助项目(61971426)