Abstract:
The detection, segmentation, classification, and crack severity indices of cracks are essential for
different kinds of surface assessment and maintenance. When analyzing the degree of cracking on
a different type of surfaces, the detection and division of cracks are critical elements to consider.
The severity of the crack influences repair and maintenance. This industry employs a variety of
image processing techniques. Traditional approaches, such as edges and deep learning algorithms,
have the potential to capture tiny amounts of information. CNNs are increasingly being used for
automated concrete crack detection. CNNs outperform traditional image processing methods. This
is particularly true for occupations requiring sophisticated identification. In open-source crack
classification, the U-Net, a semantic segmentation method based on a convolutional neural
network (CNN), beat the other strategies. The U-Net model works well for certain datasets, but
further work is required to address complicated surface crack situations. In this paper we increased
the performance of the model and achieved a higher F1 Score as compared to previous methods.
A U-Net convolutional network with a ResNet34 encoder is used in the approach. ResNet34 was
pre-trained using the ImageNet dataset. Furthermore, The use of data augmentation techniques
improved model performance during training and testing. Promising experimental results are
achieved.
