Visual Identification of Transmission Tower Faults and Semantic Attributes using Deep Learning

Abstract

A historical set of faults with accompanying images provided by network operator TrønderEnergi is studied in order to investigate the amount of variance within and between fault categories, and to suggest appropriate strategies for automating the detection of each fault category using image segmentation and object detection. This analysis is partially verified through the creation of an object detection dataset targeting a specific subset of fault categories from the historical data, from which an object detector is trained and evaluated.

Further, the effects of different types of ontological structures is studied, by comparing the performance of a flat class structure to a hierarchical class structure, and to a hierarchical class structure with semantical attributes. It is shown empirically how an ontological framework with a hierarchical structure and semantical attributes provides a more compact and natural representation of the target domain, with the best performance of the evaluated structures.