Weak Supervised Asphalt Pavement Segmentation

The labeling effort of deep learning-connected projects can introduce unexpected expenditure, especially when we are dealing with a challenging problem representation such as noisy camera images. Instead of a classical semi-supervised concept, where a small portion of labeled data have to be present, we utilized a fully automatized solution. The idea of weak supervised learning can ease the labeling trait, namely we can build a pipeline with distinctive stages of a-prior information-base automatical but not so reliable ground truth data generation, and the supervised training stage. In latter, as in a typical setup in image processing tasks, the number crunching unsupervised part is done by a Convolutional Autoencoder (AE) to create a significantly more compact representation of the pictures. Usually, AEs can be beneficial on their own standard architecture concept but they can be equipped with explicit constraints, especially during training their inner pattern extraction can be regularized by additional heads. In this paper we branch on the output of the widest latent vector representation of each pixel so the neurons inside the given layer can perform also on the segmentation-head as well as on the reconstruction. With this setup we formulated a weak supervision building block, relying only on automated labeled data generation on pavement images.