When machine learning and image classification get integrated, computers become capable of performing visual tasks that until recently could only be carried out by humans. Together, these technologies offer the potential for breakthroughs in automation, presenting new digital opportunities for companies in a variety of domains.
The possibilities are staggering, which perhaps helps to explain why image recognition has risen to become one of the top five AI industries in Europe.
This article gives some examples of current and future uses of machine learning for image classification in the healthcare, insurance, financial services, manufacturing, and automotive industries. Here, some of the most receptive adopters are enjoying success with the technology and planning to expand its use over time.
Machine Learning for Image Classification in Healthcare
Let’s begin by exploring some medical applications for image classification through machine learning. Image analysis, whether performed by a human or a machine, can literally influence life or death decisions, as doctors often depend on what they can see as much as anything else in identifying medical conditions and correct treatment for them.
In fact, according to an article published by Healthcare Informatics last year, IBM researchers say images represent 90% of all data used in the medical field. Given that computers are not prone to stress, distraction, or fatigue, it’s no great stretch of the imagination to see how imaging solutions can help medical professionals diagnose patients accurately and hence administer the most appropriate treatment.
In addition to consistent accuracy, speed is another advantage of artificial intelligence systems, such as convolutional neural networks. Once trained, they can classify and analyze visual content more quickly than humans, which means faster as well as more accurate diagnosis and treatment.
Medical Applications for Image Classification
The value of image analysis and classification software has not been lost on healthcare innovators, and computer vision is already making its way into a range of diagnostic processes and disciplines, including:
- CT, MRI, and ultrasound scanning
- Diagnosis of skin diseases
- Comparisons of new medical images with those on a patient’s historical record
- Predicting particular pathologies
Machine learning for image classification is also proving valuable in the fight against cancer, particularly for classifying breast lesions captured by ultrasound as either benign or malignant—a task traditionally falling on the shoulders, or rather the eyesight, of doctors.
A deep learning system developed by Samsung is able to lift some of that burden, and by visually identifying the presence or absence of malignancy it helps medical imaging processes to reduce the need for invasive biopsy.
Insurers Embrace Machine Learning for Image Classification
In addition to using image classification in such primary areas of healthcare as diagnosis and prognosis, it is conceivable that health insurers too can benefit from the increased speed and accuracy of the analysis. After all, earlier and more accurate diagnoses, let alone preventive measures, should result in lower treatment costs.
There are more benefits for health insurers, as the insured will be able to submit photographs of their medical records, from which image classification software will interpret the data necessary for claims processing. This will help to lessen the costs of claims administration by reducing the amount of manual processing required, as well as the likelihood of errors and misunderstandings during the adjustment process.
In other verticals too, insurers are making good use of image classification. The following examples highlight some important current and future applications of machine learning for image classification in the insurance sector.
Managing Customer Applications and Onboarding
Insurance companies will be able to process customer applications faster and with less risk of error by using image classification to extract data from an applicant’s documents. In fact, the applicants themselves can submit photos of the documents they are asked to supply. Machine learning algorithms will then process the visual data in these images to be translated and entered automatically into the insurance company’s database.
As an example of the uses of machine learning for image classification in assessing policy applications, American Family Insurance (AmFam) is exploring the technology as a way to evaluate the age and condition of roof structures. It is an important element in the risk assessments that insurers conduct before providing quotations, agreeing or declining to insure certain property.
Just as real estate insurance companies require evidence of a property’s structural condition, so life insurers must assess certain human conditions before calculating the policy price.
Body mass index (BMI) is one of the most important factors in such assessments.
AmFam believes there is the potential to assess BMI from an applicant’s selfie image captured on any mobile device. If successful, this should prove more accurate than current self-reporting methods. At the same time, it means avoiding the need for BMI to be assessed by way of a paramedical examination.
Banks are Cashing in on Image Classification Solutions
Like the insurance industry, banking institutions have seen the possibilities to save money and improve processes using fintech AI tools, including machine learning for image classification.
As evidence of the savings potential, a quick look at the reports of the COiN solution implemented by banking giant JPMorgan Chase in 2017 should be enough to get banking professionals at least a little excited, and lawyers a little worried.
A New Kind of COiN is Mint for JPMorgan Chase
COiN uses image classification along with other machine learning technologies to scrutinize and review agreements for commercial loans. Bloomberg reports that after the aptly named COiN software was implemented, it effectively wiped out the cost of 360,000 hours of manual agreement reviews.
Given that lawyers had previously carried out much of the work now performed by COiN, the financial gains for JPMorgan Chase have no doubt been phenomenal, although the bank appears not to have disclosed the actual amount of saved costs.
As with other examples of image classification covered above, COiN’s advantages exceeded initial expectations by not only saving vast amounts of time but also reducing errors, meaning loans can be now serviced more effectively.
Paper Checks Go Digital
While COiN is an outstanding example of the cost-associated advantages of image classification, it is by no means the only one. At least three banking brands in North America are using image classification as a step toward the cost and service efficiencies of “straight-through processing” (highlighted in the image below), by digitizing the process of depositing traditional checks.
Bank of America, TD Bank, and Provident Bank all accept images of checks that their customers wish to deposit. For this, customers need to photograph their checks using a smartphone camera and send the images to the bank.
Machine learning and image technologies for banking are used to classify and analyze the visual data from such images and register the deposits, saving customers from the need to present original paper checks in person at a bank branch.
This digitized process saves money for the bank by eliminating much of the manual activity involved in registering check deposits and provides customers with an improved, more convenient experience when depositing funds to their accounts.
Image Classification for Manufacturers
The manufacturing industry has already proven to be a fertile environment for machine-learning solution development, with image classification already being applied to a range of challenges, from monitoring machine wear-and-tear on the assembly line, to performing quality inspections of work-in-progress and finished products.
It is not hard to find real-world uses of image classification in manufacturing facilities, with the following examples proliferating rapidly across the industry:
- Packaging inspection: Pharmaceutical companies are using vision systems with image classification to check if medicines on the production line are whole or broken, and to check that they conform to color and size specifications.
- Barcode reading: PanelScan is a solution developed to improve the process of barcode scanning, itself an older form of image recognition. Instead of requiring humans to scan barcodes, for example, on printed circuit boards (PCBs), the PanelScan system performs the same process via fixed cameras on the assembly line.
- Quality inspection: Image recognition systems such as those supplied by Acquire Automation are able to inspect a wide variety of manufactured items for quality verification without human intervention, even when those items are cylindrical in shape and inspection must be performed from a 360-degree perspective.
Image Classification Everywhere in Automotive
From the factory floor—a bastion of automation for many years—to the slowly evolving driverless experience, machine learning, image classification, and other advanced AI concepts have reached pervasive levels in the automotive sector.
In fact, it’s hard to discuss the use of image analysis in the automotive arena without touching on a couple of the topics covered earlier in this article. So let’s begin with manufacturing, where carmakers use image classification to keep an eye on the health of their production-line robots.
Robots Watching Robots
When it comes to automotive manufacturing, there is literally no time to stand still, which is just one reason why Toyota and many other carmakers have been continuously developing and improving robotic technology on the factory floor. Yet while robots can build cars continuously 24/7/365, they do suffer occasional malfunctions or breakdowns.
As rare as those occasions may be though, they still present huge problems. According to the Robotic Industries Association, the cost of just one minute of production-line downtime for a company like General Motors costs at least $20,000. For that reason alone, predictive maintenance capabilities are high on the agenda for auto manufacturers’ IT departments.
General Motors is tackling the issue of rare but expensive downtime with the help of its cloud-based image classification software. The product is fittingly named ZDT (which stands for zero downtime) and analyzes images from cameras mounted on assembly robots, to spot signs and indications of failing robotic components.
During the pilot of the ZDT remote vision system, for which it was installed on some 7,000 robots, it successfully detected 72 instances of component failure before they resulted in unplanned line stoppages.
Motor Insurers Watching Drivers
In our earlier section on the insurance industry, we described how image classification is used to assess risk by analyzing images of buildings and people. While auto insurers have not yet harnessed the technology to the same effect, it is very possible that at some future point in time, in-car cameras and image classification software will capture and analyze images of drivers’ behaviors as they operate their vehicles, as illustrated in the image below.
The results of the analysis may be used to determine drivers’ insurability, and to calculate the rates they must pay for their cover. That said, if these expected breakthroughs eventually occur within the realm of driverless technology, driver-behavior monitoring might become obsolete before it even gets off the ground.
Yet here too image classification plays an important role—enabling driverless cars to make sense of the environment around them.
Cars Watching the Road
In fact, while an array of different sensors are used in autonomous vehicles, cameras and the images they provide are the primary means of what driverless car developers call perception.
Machine learning for image classification is vital to automobile autonomy. Driverless cars simply could not exist without the software that can learn to recognize the meaning of road signs, lane markings, and other highway features, as well as the nature of obstacles appearing in the path of the vehicle.
Image Classification in Other Industries
Healthcare, insurance, automotive, manufacturing, and financial services are among the industries in which automated image classification is most prolific. However, in reality it is likely that many more industries will be impacted by the technology as it matures.
Digital-first industries, such as ecommerce, social media, and digital marketing are already well-versed in adopting image classification. The applications include making online search more user-friendly, tagging images automatically, providing personalized recommendations, and identifying people of specific age groups for tailored marketing initiatives.
In traditional retail, some brands are exploring image classification to analyze consumers’ shopping carts as they fill them in-store. This will enable the stores to push personalized purchase recommendations to consumers’ mobile devices as they shop. Meanwhile, for high-value retail items, machine learning for image classification can be used in a similar way to augmented reality apps, enabling consumers to receive detailed information about potential purchases just by snapping a photo with their smartphones.
Tourism and Travel
Companies in the travel and transportation industries are discovering many inventive uses for image classification.
For instance, Singapore’s Changi Airport, one of the busiest transport hubs in the world, is equipping its new terminal with computer vision to leverage facial recognition during check-in, security, and departure processes, to smooth the passenger flow and prevent congestion.
Tourist venues will be able to provide visitors with information from captured images, perhaps as part of gamified guides, encouraging tourists to photograph certain highlights and features in order to receive more incentives and get interactive information about the subjects of their snaps.
Passenger and Cargo Transportation
In cargo transportation, image classification may be able to replace barcode scanning for shipment tracking, with photographed images of package labels serving as the tracking-data source.
Passenger transport operators could possibly implement facial recognition on buses, trains, or in taxis, serving both as a security measure and linking to payment systems, enabling passengers to pay their fares instantly without handing over cash or presenting debit/credit cards.
How Can Your Industry Take Advantage of Image Classification?
By analyzing images of people, places, objects, scenes, and documents, machine learning for image classification promises new levels of automation in just about every industry. If your business is not involved in the industries mentioned above (or even if it is), the examples presented here may inspire you to have your own ideas about using the technology in your company.
If that’s the case, we would love you to continue the discussion by sharing your thoughts.
Perhaps your business is already utilizing machine learning and image classification to improve service, reduce costs, or increase levels of automation and minimize errors. If so, please feel free to share your results—and your challenges—with other readers here on the Iflexion blog.