Installment 2: AI Today

Artificial Intelligence in the visual infrastructure-inspection industry

As the amount of data grows, so does the challenge of extracting value from set data. In 2020, many companies trust that Artificial Intelligence can assist in data processing. But, while AI-technologies will undoubtedly revolutionize almost every industry at some point, in the market of visual infrastructure-inspections, it is still early days.

Computer vision replacing human eyes

Artificial Intelligence, or AI, is one of the hottest buzzwords in visual data management. It seems almost destined, that tasks previously completed by experts, will be taken over by intelligent algorithms. Some even claim to have made the transition already. While it can be unclear, how far along the technology really is, the advantages of employing AI for jobs previously held by subject matter experts are crystal. Take the example of identifying and annotating faults in inspection images.

First off, eliminating human variables increases the reliability of any system. Initially, there will still be errors, but as the system improves, the accuracy will move towards 100%. At some point, we will have algorithms detecting faults with 100% reliability 100% of the time.

Secondly, experts are highly paid, and detection is a recurring task. A Machine Learning algorithm can perform more tasks in the same time span and cut costs immediately, as well as in the long term.

A third significant advantage of AI-algorithms is prominent especially for critical infrastructure fault detection. Increasing reliability and accuracy here, means fewer accidents and an overall decrease in risk to infrastructure and people.

A road paved with obstacles

Despite the broad consensus about the benefits of AI, we still have a long way to go before full automation is possible. So, what is holding progress in this area back? 

There are several challenges in the field of computer vision. At the core of them all, is the fact that we are attempting to create algorithms that can understand visual data, based on our understanding of human vision. Unfortunately, vision is a concept we do not fully understand yet, and this creates fundamental challenges.  

However, being a “refine-as-you-go” type of process, there are also more immediate problems facing the industry today.

Data

Neural networks (NNs) need massive amounts of data to become accurate enough to compete with their human counterparts, and, even though the data currently exists, it is scattered, most of it is un-labelled, and a very small part of it is being used for the purpose of training NNs. 

We need more, good, labelled data and more time to process it. The good news is, this process is happening as you read, and we will get there soon. For some specific issues and infrastructure types, there are already algorithms that work quite well. 

That leads us to the next problem.

Humans

Because so many different companies are working on algorithms, it is very hard to make them universally applicable. 

Machine learning depends on labelled data and corrective input to become accurate. With different people providing this input, the standards for it are not uniform and that is a problem. Even within the same system, the human factor can cause problems. Two people might be training an algorithm to detect corrosion on transmission poles, by annotating a large dataset, but vary in the way they create annotations or use contrasting labels. It can be hard to get around this issue, but at Scopito we are attempting to, by letting each user adapt our algorithm to suit their exact standards. Because the algorithm is already trained, it can take as little as 15 annotated images, to tech it to recognize and annotate defects in an inspection.

Despite the current challenges, the industry, has come a long way in the past few years. At the same time, fierce competition between companies, creates an environment where transparency is hard to come by. This begs the question: how far have we really come? 

What is possible today

“AI brings to the table the magic of human decision paths – they’re very complex, and unmatched by regular machine vision systems.” 

– Massimiliano Versace (Jim Vinoski, 2020)

There are many open-source AI-platforms available to companies today. These can – and often do – serve as a basis for in-house algorithms. This allows companies everywhere to adapt for their specific use-cases - if they have enough data. Some of the most popular platforms are:

• TensorFlow by Google

• Scikit-learn

• Microsoft Cognitive Toolkit 

• Theano

• Keras

• Torch

• IBM Watson

• Amazon SageMaker Neo
  
  

Other companies are creating their algorithms from scratch, and some in turn, allow users to integrate their own AI as well.

Note: There is no algorithm that can complete every task, and no one has perfected all their AI-capabilities yet. Make sure to see live proof of working AI, before committing to an it-provider, if AI is an important factor for you. 

To understand what is currently achievable in the visual infrastructure-inspection industry, we need to look at the different tasks, that an AI-based computer vision (CV) system performs. CVs can have several tasks, two of which are especially important in visual inspections: Object Detection and Instance Segmentation.

Object detection

Object detection is especially useful for identifying and classifying anomalies on inspections photos. This can work for RGB or thermal images alike, and involves the algorithm identifying, annotating, and tagging everything that resembles the types of issues, that it knows to look for. 

Object detection can be used to automate annotation of images and make datasets easily searchable. Adding decision-making algorithms to the mix, allows for an overall asset health assessment and, on the highest level, makes it possible to prioritize which assets need the most immediate attention and what actions generates the highest ROI.

Instance segmentation

Instance segmentation is a more advanced form of object detection, which includes identifying the exact pixels in an image, that belong to a certain object. The algorithm still identifies, categorizes, and annotates objects, but now on a pixel-level.

This is commonly used to identify encroaching vegetation, and can be combined with other algorithms, to determine the exact distance from a line to a branch. 

Combined with decision-making algorithms, it can be used to provide a list of all areas of line, with encroachment issues.

Other common uses of computer vision in the visual infrastructure-inspection industry are 3D modelling and asset visualizations.

This is what Artificial Intelligence can do in 2020; it can detect objects with high accuracy, if it has enough data, and it can make decisions based on what is found, to help us manage our assets’ health. 

The current uses for AI are amazing. They provide a sea of untapped potential to companies everywhere. That being said, the possible use-cases for Artificial Intelligence in the industrial inspection industry are virtually endless. In the next article of this series, we explore the future of AI in the visual infrastructure inspection industry.

Article written by: Scopito