Using Deep Learning in Computer Vision

Recent advances in deep learning have boosted the already rapidly developing field of computer vision. In particular, deep learning methods have become a central part of a wide range of computer vision applications, including image classification, localization and detection.

In this presentation we will show how we are combining traditional computer vision methods and state-of-the-art visual detection methods based on deep learning to create contextualized 3D models from images. In particular, we use photogrammetry to create 3D models of environments based on overlapping images. Further, we use various feature detection algorithms based on convolutional neural networks (CNN) to extract relevant information from those images. This includes detection and localization of certain objects in the images, which are later mapped over and highlighted in the 3D model, resulting in a contextualized 3D model.

An example application of the pipeline outlined above, is to detect and localize faulty electrical cables on oil rigs where electrical cables and highly combustible gases are separated by thin barriers. To achieve this, we have designed a new CNN-based algorithm capable of detecting electrical faults in images of cables. Moreover, we use an attention mechanism in our algorithm to not only detect, but also locate the position of the fault in the image. Whereas previous studies have used attention to locate objects in an image, there seems to be no research using it in combination with a classification network, to guide the model towards the correct place in the image, during training.

Speakers: Cim Stordal og Milad Mobarhan from Cognite.