|
Abstract:
|
Typical computer vision tasks, such as image classification, usually require a large sample of training data to make accurate classifications. In some applications the collection of images is costly, time-consuming, or no longer possible, resulting in limited image data available for model training and development. In these instances, image classification models can perform poorly and achieve low classification accuracies. Generative models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), allow for the creation of synthetic images based on a (potentially small) set of training images. Using these models small training datasets can be inflated using simulated images. This study applies this approach in a solar energy setting to improve the classification accuracy of photovoltaic (PV) system faults using thermal images of PV modules in South Africa – where limited training data is available. The results show that the use of generative models in this way improves PV fault classification accuracies. As such, generative models can potentially provide a means to perform specialised image classification tasks even when training data are limited.
|
