Thesis (MSc) - Cyprus International University. Institute of Graduate Studies and Research Information Systems Engineering Department

Includes bibliography (sheets 81-87)

ABSTRACT
Concrete is one of the most commonly used material in civil engineering and
construction. It is mainly used in the construction of buildings, bridges, pavements etc.
These concrete infrastructures are prone to cracks and these cracks are caused by
various factors. Early detection of concrete cracks guarantees the safety of concrete
infrastructures. The main focus of this work is to develop an automated deep learning
model for detecting concrete cracks. This research uses the Concrete Crack Image
Classification (CCIC) dataset having two classes. The dataset is balanced containing
20,000 images with cracks (positive class) and 20,000 images without cracks (negative
class).
In this work, the effect of converting images to grayscale and fine tuning hyper
parameters (batch size and Adam LR) was investigated by performing several
experiments. Converting the images to grayscale resulted in simpler models with
reduced parameters due to the one channel of grayscale images as compared to the
three channels of RGB images. Two different architectures of CNN models, CNN 1
and CNN 2 were built and trained on RGB and grayscale images. The experimental
results revealed that there is no much significant difference in the performance of the
RGB and grayscale CNN models. This means that concrete crack detection relies more
on structural patterns rather than color information. Although, converting the images
to grayscale allows the model to focus more on learning the crack features.
The models were further tested on an external data different from the development
data. In this experiment, the CNN 2 model using grayscale images (batch size of 32
and Adam learning rate of 0.001) outperformed the other models and is chosen as the
proposed model. The proposed model achieved a little change decrease in the
difference in performance between the development data and external data.
In addition, the proposed model’s generalization ability was further validated using 10
fold cross validation method. This was done to ensure that the model is not overfitting.
The proposed model achieved a mean accuracy of 99.74%, mean precision of 99.83%,
mean recall of 99.65%, mean F1-score of 99.74%, mean specificity of 99.83%, and
mean AUC ROC of 0.9974. These performances were further compared with seven
previous concrete crack detection studies that have used the CCIC dataset. From the
comparative analysis, the proposed model performed better than the solutions of six
of these studies in all the performance metrics compared.
Keywords: Computer Vision, Concrete Crack Detection, Convolutional Neural
Network, Deep Learning