An improved DeepLabv3 + railway track extraction algorithm based on densely connected and attention mechanisms

Abstract

Abstract The railway track extraction using unmanned aerial vehicle (UAV) aerial images suffers from issues such as low extraction accuracy and high time consumption. In response to these problems, this paper presents a lightweight algorithm DA-DeepLabv3 + based on densely connected and attention mechanisms. Firstly, the lightweight MobileNetV2 network is employed to replace the Xception feature extraction network, thereby reducing the number of model parameters. Secondly, the receptive field is enlarged by cascading atrous convolutions with different dilation rates in the ASPP (atrous spatial pyramid pooling) module, and other feature maps are concatenated using the multi-scale attention module to enhance the extraction accuracy of the model. Finally, a multi-level upsampling module is designed to enhance the accuracy of boundary contour extraction. Furthermore, a dedicated dataset for railway track segmentation was established to train and evaluate the proposed method. The experimental results indicate that DA-DeepLabv3 + demonstrates significant improvement on the railway track segmentation dataset as well as the DeepGlobe dataset. It achieves mIoU scores of 87.52% and 85.01%, along with accuracy rates of 97.59% and 94.84%, respectively. Compared to classical semantic segmentation networks such as U-Net and DeepLabv3 + , DA-DeepLabv3 + achieves higher extraction accuracy and shorter running time.

Keywords

• Deep learning
• Semantic segmentation
• Attention mechanism
• Railway extraction
• UAV aerial imagery