Domain adversarial training for classification of cracking in images of concrete surfaces

The development of automatic methods to recognize cracks in surfaces of concrete has been under focus in recent years, firstly through computer vision methods and more recently focusing on convolutional neural networks that are delivering promising results. Challenges are still persisting in crack recognition, namely due to the confusion added by the myriad of elements commonly found on concrete surfaces. The robustness of these methods would deal with these elements if access to correspondingly heterogeneous datasets was possible. Even so, this would be a cumbersome methodology, since training would be needed for each particular case and models would be case dependent. Thus, efforts from the scientific community are focusing on generalizing neural network models to achieve high performance in images from different domains, slightly different from those in which they were effectively trained. The generalization of networks can be achieved by domain adaptation techniques at the training stage. Domain adaptation enables finding a feature space in which features from both domains are invariant, and thus, classes become separable. The work presented here proposes the DA-Crack method, which is a domain adversarial training method, to generalize a neural network for recognizing cracks in images of concrete surfaces. The domain adversarial method uses a convolutional extractor followed by a classifier and a discriminator, and relies on two datasets: a source labeled dataset and a target unlabeled small dataset. The classifier is responsible for the classification of images randomly chosen, while the discriminator is dedicated to uncovering to which dataset each image belongs. Backpropagation from the discriminator reverses the gradient used to update the extractor. This enables fighting the convergence promoted by the updating backpropagated from the classifier, and thus generalizing the extractor enabling it for crack recognition of images from both source and target datasets. Results show that the DA-Crack training method improved accuracy in crack classification of images from the target dataset in 54 percentage points, while accuracy on the source dataset remains unaffected.