Brief Review — ResUNet: Road Extraction by Deep Residual U-Net
ResUNet, U-Net With Residual Paths
2 min readApr 2, 2023
Road Extraction by Deep Residual U-Net,
ResUNet, by Beihang University,
2018 IEEE Geoscience and Remote Sensing Letters, Over 1500 Citations (Sik-Ho Tsang @ Medium)Semantic Segmentation
2014 … 2021 [PVT, PVTv1] [SETR] [Trans10K-v2, Trans2Seg] [Copy-Paste] 2022 [PVTv2] [YOLACT++]
- A semantic segmentation neural network, ResUNet, is proposed, which combines the strengths of residual learning and U-Net, is proposed for road area extraction.
Outline
- ResUNet
- Results
1. ResUNet
- In this paper, residual block or short skip connection (blue dahed lines) is applied instead of just using plain convolutions as in U-Net.
- 7-level architecture of deep ResUnet is used.
- MSE is used as loss function:
- where N is the number of the training samples.
- Overlap strategy, as in U-Net, is also used here.
2. Results
- The break-even point is defined as the point on the relaxed precision-recall curve, where its precision value equals its recall value.
ResUNet performs better than all other three approaches in terms of relaxed precision and recall.
- (a) Input image. (b) Ground truth. (c) Saito et al. [5]. (d) U-Net. (e) Proposed ResUNet.
Even the runway has very similar features to a highway, the proposed method can successfully segment side road from the runway.
