Data augmentation using generative adversarial networks for electrical insulator anomaly detection

Abstract

Electricity has been an essential part of our life. Insulators, which are widely used for electricity transmission, are prone to be damaged and need constant maintenance. Traditionally, the inspection job is time-consuming and dangerous as workers would have to climb up the electricity tower. Deep learning has offered a safe and quick way to inspections. About 3000 insulators images are taken from different angles using a drone. Due to great difference in number of good and damaged insulator, directly training a classifier on the imbalanced data lead to low recall value on the damaged insulators. Generative adversarial networks (GANs) were introduced as a novel way to augment data. However, traditional GANs are either incapable of generating high quality images or fail to generate minority class images when minority class examples are far less. In this study, a novel GAN model, Balancing and Progressive GANs (BPGANs), was proposed for effectively making use of all classes information and generating high quality minority images at the same time. Results show that PGANs, StyleGANs, and BPGANs were able to generate high-resolution images and improve classification performance. PGANs achieved the better results than BPGANs. This may be because BPGANs only provides 2 additional latent codes since it is a binary classification, having little effect on generating desired images. BPGANs seemed to have difficulties generating class-specific images, which might be because that the classification loss is too little compared to the source loss and optimization was more focused to optimize the source loss. This indicates that learning representations of data progressively from low resolution to high resolution is an effective approach, however, embedding class label information in the fashion of AC-GANs and BGANs might not be appropriate for augmenting binary class data sets.