Transformed-based foundational models in Computer Vision: an industrial use case

Abstract

The adoption of deep learning models in industries for computer vision tasks is often hindered by the need to collect an adequate number of images to build a dataset sufficiently representative of the problem to be addressed. The common approach for handling this issue involves starting from a pre-trained model and finetuning it on a domain-specific dataset, drastically reducing the required number of images and, at the same time, delivering acceptable performances. Recent advances in neural network architectures such as transformers and self-supervised training techniques have contributed to the development of a particular type of deep learning model, namely, foundational models, that can be easily adapted to a wide range of downstream tasks. In this work, the performances of two foundational models were compared in a real industrial use case requiring the detection and classification of welding defects for small components. The application of these models as the backbone for a single classification linear layer trained on a small, curated dataset of 80 images, achieved an accuracy of around 95% on the test and validation sets with little implementation effort.