In this work we show how Group Equivariant Convolutional Neural Networks use subsampling to learn to break equivariance to their symmetries. We focus on the 2D roto-translation group and investigate the impact of broken equivariance on network performance. We show that changing the input dimension of a network by as little as a single pixel can be enough for commonly used architectures to become approximately equivariant, rather than exactly. We investigate the impact of networks not being exactly equivariant and find that approximately equivariant networks generalise significantly worse to unseen symmetries compared to their exactly equivariant counterparts. However, when the symmetries in the training data are not identical to the symmetries of the network, we find that approximately equivariant networks are able to relax their own equivariant constraints, causing them to match or outperform exactly equivariant networks on common benchmark datasets.

Convolutional Neural Networks are particularly vulnerable to attacks that manipulate their data, which are usually called adversarial attacks. In this paper, a method of filtering images using the Fast Fourier Transform is explored, along with its potential to be used as a defense mechanism to such attacks. The main contribution that differs from other methods that use the Fourier Transform as a filtering element in neural networks is the use of labeled data to determine how to filter the pictures. This paper concludes that, while the filtering proposed is hardly better than a simple low-pass filter, it still manages to improve resistance to adversarial attacks with a minimal drop in the standard accuracy of the network.

This paper shows how the current state of the art in image classification performs on LEGO bricks. Currently the standard image classification models with deep learning are single label image classifiers. In this paper we will convert them to work on multi-label images and subsequently evaluate how well they perform. We show how well the classifiers will work on three different types of datasets. Experiments will be conducted on these three types of datasets to compare the performance of three different multi-label image classifiers. The main research question accompanying this paper is ``How well does the state of the art in image classification work on LEGO bricks?''. Three subquestions are set up to answer this question. The first will regard the existence of the image classifiers. The second subquestion will regard how big the influence is of real life aspects, such as deterioration of the LEGO bricks. The final subquestion will be about the performance on the datasets. After answering these questions and conducting the experiments, we came to the conclusion that the ResNext model performed the best on almost all of the categories. Based on the numbers of the results we can also conclude that the models should perform well with multi-label images of LEGO bricks.

Data collection by means of crowdsourcing can be costly or produce inaccurate results. Methods have been proposed for solving these problems. However, it remains unclear what methods work best in scenarios with multiple similar objects of interest present in the same image, which is important for training computer vision with applications such as automatic quality control in factories. We researched which parameters are important to optimize, which methods are worth considering and what those selected methods score with regard to the parameters cost and quality. This was done through a literature review and substantiated by an experimental crowdsourcing campaign that focused on the annotation of Legos in images. It was found that the parameters to optimize were cost, optimized by reducing the time workers spent on tasks, and quality, optimized by improving the mean intersection over union value of the annotations. We concluded that majority vote, rejecting workers, majority vote adjusted to be resistant to outliers, rejecting workers with the same adjustments and decomposing tasks were the most promising methods. From our experiment we concluded that a clear trade-off exists between cost and quality. The adjusted rejecting workers method, that uses worker credibility, showed to have the highest mean quality. While the method that decomposed the components of the task and distributed them was the cheapest method to use overall and also best when looking at mean quality over cost, it was worse quality wise. These results were similar to the expected performance of the methods. From this we concluded that the best method for crowdsourcing is dependent on the error tolerance of the computer vision model that will be used and the budget available.