Improved Pixel-Level Pavement-Defect Segmentation Using a Deep Autoencoder 📄 [1]

📜 Article can be downloaded (open access) from:
https://www.mdpi.com/1424-8220/20/9/2557

If you find code or ideas useful, please cite [1],[2]

🔥 Information about training, prediction and computational performance can be found in the article 📄.

✅ Solutions:

• UNet
• ResUNet
• ResUNet with Atrous Spatial Pyramid Pooling
• ResUNet with Atrous Spatial Pyramid Pooling ("Waterfall"[3] connection)
• ResUNet with Atrous Spatial Pyramid Pooling and Attention Gates
• ResUNet with Atrous Spatial Pyramid Pooling ("Waterfall"[3] connection) and Attention Gates

⚠️ It is not all! Feel free to make your own configuration using neural network block, defined in 'models/layers.py' 🐍 file . You may find even more architectural solutions in the code than we mentioned ☝️ 👀 ☝️.

Baseline model:

Model induced with residual connections, ASPP, AG:

Few results with different architectures:

Third-party library priorities:

• Tensorflow (with Keras frontend)
• OpenCV
• Many nifty things related to Python programming language

Usage:

Everything is straight-forward. Check comments in the code 👀

• train.py - train
• predict.py - predict
• predict_by_patches.py - predict a big image by cropping it into regions and joining them after

Rendered video results:

---

CrackForest links:

UNet, ResUNet, ResUNet+ASPP, ResUNet+ASPP+AG, ResUNet+ASPP_WF, ResUNet+ASPP_WF+AG

---

GAPs384 links:

UNet, ResUNet, ResUNet+ASPP, ResUNet+ASPP+AG, ResUNet+ASPP_WF, ResUNet+ASPP_WF+AG

---

Crack500 links:

UNet, ResUNet, ResUNet+ASPP, ResUNet+ASPP+AG, ResUNet+ASPP_WF, ResUNet+ASPP_WF+AG

---

References

[1] Augustauskas, R.; Lipnickas, A. Improved Pixel-Level Pavement-Defect Segmentation Using a Deep Autoencoder. Sensors 2020, 20, 2557.
[2] Augustauskas, R.; Lipnickas, A. Pixel-wise Road Pavement Defects Detection Using U-Net Deep Neural Network. In Proceedings of the 2019 10th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS), Metz, France, 18–21 September 2019; IEEE: Metz, France, 2019; pp. 468–472
[3] Artacho, B.; Savakis, A. Waterfall Atrous Spatial Pooling Architecture for Efficient Semantic Segmentation. Sensors 2019, 19, 5361.