|
A Deep Learning Approach based on Specialized Convolutional Blocks in Urban Road Extraction
|
Mohammad Erfan Omati   , Fatemeh Tabib Mahmoudi * |
| Shahid Rajaee Teacher Training University |
|
|
Abstract: (560 Views) |
| Road extraction from high-resolution remote sensing images has been used in a wide range of applications such as traffic management, route planning, and road navigation. Due to their long length and small width, as well as shadows caused by vegetation and buildings, the detection of the roads challenging. As the roads In an area are of different types such as being near short passages, highways and motorways, we face some difficulties in automatic classifying and recognizing different kinds of roads. In order to improve the reliability and accuracy of extraction of the roads with shorter lengths when there are roads of different sizes; a neural network model is proposed in this paper that achieves pixel-accurate segmentation. The proposed network directly processes the input image and uses four specialized convolutional blocks (SCB) during down-sampling which is complemented by a shallow sampling approach to generate a binary mask for the road class. As the common semantic segmentation networks are deep and have various teachable parameters, the proposed network in this research uses shallow sampling which leads to lessen the network depth and as a result the number of the parameters decreases. The performance of the proposed model in this research was evaluated using the Massachusetts dataset, and the evaluation results clearly show the superior performance of the proposed model compared to the other neural networks with fewer parameters. Compared to the other neural networks such as DEEPLAB3+, U_NET and D_LINKNET, the proposed model was able to improve the IOU and F-Score indices in Massachusetts dataset by 1.98 and 3.03, respectively. |
|
| Keywords: Semantic Segmentation, road extraction, Deep Learning, specialized convolutional blocks |
|
|
Full-Text [PDF 868 kb]
(150 Downloads)
|
Type of Study: Research |
Subject:
RS
Received: 2024/01/6 | Accepted: 2024/06/12 | Published: 2024/06/20
|
|
|
|
|
|
|
| Send email to the article author |
|
|
