Optimal Selection of Data Augmentation Transformations

To achieve useful results in training deep neural network, one typically needs huge number of labelled images. In most practical cases, it is often time consuming, expensive and sometimes even near impossible to get the required numbers of labelled images. Data augmentation consists in a set of image processing techniques that can be applied to an image to generate another slightly different image. By doing so, one can efficiently increase the number of different images to feed the training phase of a neural network. This works well until the generated images are too different from the original images. When this happens, the performance of the trained neural network is negatively impacted. The objective of this project is to learn the combination of image processing techniques that can be applied to a given set of images that will yield the best possible trained neural network. The complexity lies not only in the construction of the image processing pipeline but also in the configuration of each image processing algorithms inside the aforementioned pipeline.

Faculty Supervisor:

Marco Pedersoli

Student:

Antoine Harlé

Partner:

Teledyne DALSA

Discipline:

Engineering - other

Sector:

Manufacturing

University:

École de technologie supérieure