For the aerospace industry, visual surface inspection is an important task in quality control. If automated, there can be a significant reduction in manual work and error rate. For the intention behind this thesis, GKN Aerospace Sweden AB is looking to explore automated visual inspection through Machine Learning. By having the investigated surfaces photographed, the images can be fed to a computer where they are classified as either defective or non-defective. The problem is quite specific as there are limitations to the amount of available data, which is something generally counterproductive to achieving good performance with Machine Learning models. The approach is therefore focused towards utilizing methods that maximizes performance with this limitation in mind. One of the used techniques, data augmentation, concerns expanding the dataset through generating synthetic samples. The thesis covers four different ways of data augmentation, from which all proved to be beneficial (for the given amounts of expansion). Furthermore, the thesis work also investigates the use of transfer learning on common Convolutional Neural Network architectures. This means leveraging previously extracted knowledge from a different classification problem to the visual inspection problem at GKN. Transfer learning was shown to be very useful. In addition, by varying the degree to which the transferred knowledge is allowed to be modified, additional performance increases were observed.Regarding the choice of Convolutional Neural Network architecture there were tendencies for some models to perform better than others, but with no remarkable performance differences. The results showed very high accuracies for many of the suggested models, proving that Machine Learning is indeed very useful for surface inspection and can be advantageous even in cases of very limited datasets.