Interpreting visual data is something so simple for humans, we don’t realise we are doing it. However, for computers, identifying the contents of an image can be a huge challenge. Image recognition is the technology tasked with giving computers the ability to identify objects, places, people, actions, and other variables within images. It is about enabling computers to see and react to images as humans do. Developing the technology is imperative as many emerging technologies, such as autonomous vehicles and facial recognition, depend on it.
MarketsandMarkets estimates that the image recognition market will grow from $16 billion in 2016 to $39 billion by 2021. This predicted increase suggests there is considerable potential for technology to impact the market over the next few years.
Image recognition requires a camera to capture images which can then be converted into data. LDV Capital (2017) predicts that the number of cameras integrated into products, such as in cars, wearable devices, and even kitchen appliances, will increase by 220% in the next five years. This rate of growth is fundamental if we are to see the image recognition market grow at the rate that has been predicted.
Image recognition requires data in order to learn and improve. Computers can be taught the process of identifying the contents of an image, but they need an immense amount of labelled data to be able to do it quickly and accurately. Access to open source datasets is the key to this. The more data, or the more preloaded images, the computer has access to, the better chance the computer has at identifying the contents of an image. For example, an orange can be identified as an orange, because the computer already has access to how an orange actually looks. Databases, such as ImageNet, one of the largest open source database, are available for this purpose. ImageNet has over 14 million images that can be used by companies to feed and improve their algorithms.
Google and Facebook, and several other large technology companies, have their own methods of gathering the required data. They encourage users to upload and label photos to online drives.The tech matches the makeup of the images with the individual words, identifies what each image portrays, and filters the information back into their systems. This is what powers Google’s image object detection system: TensorFlow Object Detection API.
The most advanced algorithms for analyzing imagery are based on an artificial neural network, which is a “transient state that allows the machine to learn in a more sophisticated manner” The sheer quantity of data that the algorithms have access to means that the accuracy of the best artificial neural networks are arguably better than human level at deciphering the contents of an image. They perform better than humans when it comes to placing objects into specific classes, such as breeds of dog. They struggle, however, when images are placed behind colored filters or when reading human emotions–issues that humans have little problem overcoming. The technology is certainly here, it’s just a matter of harnessing it and effectively implementing it into specific industries.
E-commerce is one industry that can benefit hugely from image recognition technology. Visual merchandising has been fundamental to the industry, given that sales have always been driven by aesthetics. Nowadays, e-commerce allows shopping to be done anywhere, and, combined with image recognition technology, new forms of merchandising are appearing on the market. eBay launched a visual search tool last year that allows shoppers to search for items based on images they have taken themselves. Amazon has recently released its Echo Look, a style assistant that takes photos of your outfits and suggests fashion recommendations. Visual search technology is not new, but it could transform the way we shop completely if comprehensively adopted by the e-commerce industry.
Image recognition can massively shake up traditional practices in the insurance industry, particularly in the claims process. Insurers spend hours identifying whether claims are valid. Assessing physical damage is a visual task, for example, which has always required a claims assessor to check over the damaged object – a lengthy and costly process. The use of image recognition technology allows images of damaged objects uploaded by policyholders to be assessed straight away. The legitimacy of claims can be determined automatically, and information regarding repair costs and logistics can be subsequently provided to both the insurer and the policyholder–a process that takes very little time and a fraction of the cost of the traditional claims process method. This can be applied across multiple insurance types, from motor to health to home insurance.
ViewSpection, a Hartford InsurTech Hub startup, provides an insurance inspection app for residential and commercial properties. The app allows the policyholder to undertake their own inspections by completing a photo tour of their property – with guidance provided by the app – resulting in a more frictionless process. For insurers, property data on a policy can be gathered much faster through Viewspection’s app.
“It can take up to 40 days for a traditional inspection to be delivered… ViewSpection is delivered on day one,” says Jay Kramer, COO, ViewSpection. “Insurers can use image recognition to filter and flag the reports that an underwriter needs to look at.”
Image recognition technology has advanced to a stage where it can be used to great benefit across multiple industries. It is already making big strides in e-commerce, and the automotive industry is relying on image recognition as part of the technology being used in autonomous vehicles. The insurance industry must embrace it to satisfy consumer demands and, almost more importantly, shake off its reputation of being lengthy, arduous, and old-fashioned.