Road Segmentation Algorithm Based on Simplified Mask R-CNN for GF-3 SAR Images

Road information is a major part of feature information of the Gaofen-3 (GF-3) synthetic aperture radar (SAR) images, which has a great effect on path planning and intelligent transportation. Because of the simple road features, the common network models can segment the road targets accurately, but they usually take more time and occupy more memory. To solve the problem, a simplified Mask R-CNN method is proposed. On the basis of ensuring the accuracy of road segmentation, the original Mask R-CNN network structure, parameter scale and other aspects are simplified to improve the memory occupation and execution time. The proposed algorithm is compared with Mask R-CNN and PSPnet segmentation algorithms with GF-3 SAR image road dataset. The experimental results indicate that the proposed algorithm has a good segmentation effect, and the PA (pixel accuracy) value is more than 90%. At the same time, the size of the model is reduced to 16MB, and the running speed is also greatly improved.