Vision-based AI for infrastructure inspection support

Modern-day societies are fundamentally dependent on electricity. This poses severe requirements on maintaining the reliability and availability of electricity supply. To prevent outages and to maintain secure and reliable supply, electric utilities are required to perform regular inspections on their grids. These inspections have been typically carried out using a airborne surveys via low-flying helicopters and field surveys via foot patrols. Some utilities document only potential defects and anomalies, while others take pictures of the whole grid including pictures of conductors, power line components and surrounding objects (e.g. vegetation). The collected images are then manually inspected to identify defects. Recently, breakthroughs in Artificial Intelligence (AI) and more specifically in Deep Learning (DL) with Convolutional Neural Networks (CNNs), have revolutionized the field of computer vision and opened up new opportunities for automating the data analysis in vision-based inspections. eSmart Systems has been at the forefront of the development and application of CNN-based solutions for inspection support. Inspired by these achievements, in this project, we wish to push the technology further and attempt to solve some of the main limitations of current CNN-based solutions for object detection, image classification, anomaly detection, etc. The primary objective of the project is to improve the applicability of vision-based AI (primarily CNNs) to cases where the amount of training data is limited. This is often the case when AI needs to recognize rare objects or defects. CNNs typically require thousands of examples of each object type or class to train properly and fail to perform adequately in many infrastructure inspection tasks.

Modern-day societies are fundamentally dependent on electricity. This poses severe requirements on maintaining the reliability and availability of electricity supply. To prevent outages and to maintain secure and reliable supply, electric utilities are required to perform regular inspections on their grids. These inspections have been typically carried out using a airborne surveys via low-flying helicopters and field surveys via foot patrols. Some utilities document only potential defects and anomalies, while others take pictures of the whole grid including pictures of conductors, power line components and surrounding objects (e.g. vegetation). The collected images are then manually inspected to identify defects. Recently, breakthroughs in Artificial Intelligence (AI) and more specifically in Deep Learning (DL) with Convolutional Neural Networks (CNNs), have revolutionized the field of computer vision and opened up new opportunities for automating the data analysis in vision-based inspections. eSmart Systems has been at the forefront of the development and application of CNN-based solutions for inspection support. Inspired by these achievements, in this project, we wish to push the technology further and attempt to solve some of the main limitations of current CNN-based solutions for object detection and image classification. The primary objective of the project is to improve the applicability of vision-based AI (primarily CNNs) to cases where the amount of training data is limited. This is often the case when AI needs to recognize rare objects or defects. CNNs typically require thousands of examples of each object type or class to train properly and fail to perform adequately in many infrastructure inspection tasks.