Source-linked AI summary
Road Extraction by Deep Residual U-Net
Zhengxin Zhang, Qingjie Liu, Yunhong Wang
TL;DR
Road extraction from high-resolution remote-sensing imagery remains difficult because noise, occlusions, and complex backgrounds hinder accurate labeling. ResUnet combines residual learning with U-Net architecture and rich skip connections for semantic segmentation. On the Massachusetts roads dataset, it outperforms the compared methods while using one-quarter of U-Net’s parameters.
Problem
High-resolution remote-sensing road extraction remains challenging because imagery contains noise, occlusions, and complex backgrounds.
Method
ResUnet combines residual units with U-Net architecture and skip connections within and between encoding and decoding paths.
Results
ResUnet outperforms U-Net and two other state-of-the-art deep-learning road-extraction methods on the Massachusetts roads test set.
Takeaways & Limitations
The proposed design achieves better road-extraction performance with fewer parameters than U-Net.
Abstract
from arXiv · showhide
Road extraction from aerial images has been a hot research topic in the field of remote sensing image analysis. In this letter, a semantic segmentation neural network which combines the strengths of residual learning and U-Net is proposed for road area extraction. The network is built with residual units and has similar architecture to that of U-Net. The benefits of this model is two-fold: first, residual units ease training of deep networks. Second, the rich skip connections within the network could facilitate information propagation, allowing us to design networks with fewer parameters however better performance. We test our network on a public road dataset and compare it with U-Net and other two state of the art deep learning based road extraction methods. The proposed approach outperforms all the comparing methods, which demonstrates its superiority over recently developed state of the arts.
I. INTRODUCTION
Road area extraction remains difficult because high-resolution remote-sensing imagery contains noise, occlusions, and complex backgrounds. The paper proposes ResUnet, combining residual units with U-Net connections while removing cropping.
- High-resolution remote-sensing imagery makes road extraction challenging because of noise, occlusions, and complex backgrounds.
- Road area extraction produces pixel-level road labels, whereas centerline extraction detects road skeletons.
- Deep learning methods have achieved state-of-the-art performance across computer-vision and remote-sensing tasks.
- ResUnet combines residual learning and U-Net architecture, using residual units instead of U-Net's plain neural units.
- Removing cropping makes the proposed network more elegant and is reported to improve performance.
A. Deep ResUnet
U-Net addresses deep-segmentation training difficulty by propagating low-level features to corresponding high levels while retaining semantic information.
- A. Deep ResUnet: U-Net uses low-level details while retaining high-level semantic information to obtain finer segmentation results.
- A. Deep ResUnet: Copying low-level features to corresponding high levels creates a path for information propagation between network levels.
- A. Deep ResUnet: Training deep segmentation networks is particularly difficult when only limited training samples are available.
2) Residual unit:
Residual units address the training and degradation difficulties associated with deeper networks through identity mappings and residual functions.
- 2) Residual unit:: Deeper networks can improve performance but may hamper training and cause degradation problems.
- 2) Residual unit:: A residual unit maps its input through an identity path and a residual function before activation.
- 2) Residual unit:: In the proposed network, full pre-activation residual units are used to build deep ResUnet.
3) Deep ResUnet:
Deep ResUnet is a seven-level semantic-segmentation architecture for road extraction, combining residual units with encoding–decoding paths and skip connections. Its design removes U-Net cropping and uses fewer convolutional layers.
- 3) Deep ResUnet:: Deep ResUnet uses a seven-level architecture with encoding, bridge, and decoding parts for pixel-wise semantic segmentation.
- 3) Deep ResUnet:: All three network parts use residual units containing two 3 × 3 convolution blocks and an identity mapping.
- 3) Deep ResUnet:: The decoding path upsamples lower-level features and concatenates them with corresponding encoding-path features.
- 3) Deep ResUnet:: The network has 15 convolutional layers versus 23 in U-Net, and U-Net's cropping operation is unnecessary.
B. Loss function
The network estimates parameters by minimizing the loss between predicted and ground-truth segmentations, using mean squared error optimized with stochastic gradient descent.
- Training estimates network parameters by minimizing the loss between generated and ground-truth segmentations.
- Mean squared error is used as the loss function for segmentation training.
- Stochastic gradient descent optimizes the network, although other differentiable loss functions can also be used.U-Net is cited as using pixel-wise cross entropy instead.
- The paper describes U-Net paths using encoding and decoding terminology for feature extraction and up-convolution stages.
C. Result refinement
The refinement procedure addresses boundary inaccuracies by processing overlapping sub-images and stitching their predictions into large-image segmentation results.
- 224×224 inputs and outputs have lower accuracy near boundaries because convolution uses zero padding.
- An overlap strategy uses sub-images with overlap o = 14 to refine segmentation of large images.
- The refined large-image results are obtained by stitching the overlapping sub-image predictions.
- The method is evaluated on the Massachusetts roads dataset against Mnih’s method, Saito’s method, and U-Net.
A. Dataset
The Massachusetts roads dataset provides diverse aerial imagery for training, validation, testing, and pixel-level road extraction evaluation using relaxed precision and recall.
- The dataset contains 1171 images: 1108 for training, 14 for validation, and 49 for testing.
- Each image is 1500 × 1500 pixels with a resolution of 1.2 meter per pixel.
- The dataset covers roughly 500 km2 across urban, suburban, and rural areas with varied ground objects.
- Precision measures correctly predicted road pixels among predicted roads, while recall measures labeled road pixels correctly predicted.
- Relaxed precision and recall allow matches within ρ pixels, with ρ = 3 in this experiment.
D. Comparisons
On the Massachusetts roads test set, the proposed method is compared with deep-learning road extraction baselines and shows stronger relaxed precision and recall with a smaller model.
- The proposed method performs better than all three compared approaches in relaxed precision and recall.
- 7.8M parameters versus U-Net’s 30.6M accompanies promising improvement on road extraction.
- Table II compares the methods using break-even point, where higher values indicate better precision and recall.
- Example results show cleaner, less noisy segmentation and better separation of two-lane roads and intersections than competing methods.
- Context information helps distinguish roads from similar objects and supports labeling roads obscured by trees.
- The method misses roads in parking lots when those roads are mostly unlabeled and therefore treated as background.
IV. CONCLUSION
ResUnet combines residual learning with U-Net to extract roads from high-resolution remote sensing images. Its skip connections facilitate information propagation and training while supporting a compact, high-performing network.
- ResUnet combines residual learning and U-Net for road extraction from high-resolution remote sensing images.
- Skip connections within residual units and between encoding and decoding paths facilitate information propagation during forward and backward computations.
- The resulting architecture eases training and enables a simple yet powerful neural network.
- ResUnet outperforms U-Net with only 1/4 of its parameters and also surpasses two other state-of-the-art deep learning road extraction methods.