Crack detection, classification, and characterization are key components of automatic structural health monitoring systems. Convolution based encoder-decoder deep learning architecture have played a significant role in developing crack segmentation models possessing limitations in capturing the global context of the image. To overcome the stated limitation, in the present study we propose a novel Multi-Scale Attention based Efficient U-Net which effectively tries to solve this limitation. The proposed method achieved an F1 Score of 0.775, an IoU of 0.663 and an accuracy of 97.3% on Crack500 dataset improving upon the current state-of-the-art models.

python train.py --path_imgs "path_to_image_folder" --path_masks "path_to_mask_folder" --out_path "out_path_to_store_best_model_and_logs"

python evaluate.py --path_imgs "path_to_image_folder" --path_masks "path_to_mask_folder" --model_path "path_to_saved_model" --result_path "path_to_save_results_from_test" --plot_path "path_to_store_plots"

Crack500 Dataset: This dataset includes 500 images of pavement cracks with a resolution of 2000 x 1500 pixels collected at Temple University campus using a smartphone by [1]. Each image is annotated on a pixel level. Images this large won’t fit on GPU memory; therefore, [1] patched each image into 16 smaller images. The final resolution of each image was 640x320, and the dataset has 1896 images in training set, 348 images in validation set, and 1124 images in testing set. The comparisons with state-of-the-art models were made on the results from the testing set

• Colab Notebook link: Colab Notebook
• Best Model and Logs: Best Model and Logs

• Dockerise code

[1] F. Yang, L. Zhang, S. Yu, D. Prokhorov, X. Mei, and H. Ling, “Feature pyramid and hierarchical boosting network for pavement crack detection,” IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 4, pp. 1525–1535, 2020.

[2] Lyu, C., Hu, G. & Wang, D. Attention to fine-grained information: hierarchical multi-scale network for retinal vessel segmentation. Vis Comput 38, 345–355 (2022). https://doi.org/10.1007/s00371-020-02018-w