The rise of computer vision

Mechanics, use cases, real world successes

See the world with clarity COMPUTER VISION USES

Financial services: Fraud detection, credit

Introducing the exponential business For CIOs, data scientists and analytics leaders
application, automotive damage estimation
whose organizations are considering the potential of
value of computer vision
computer vision, note what a significant growth area Government: Smart cities, security intelligence,
Who would have thought that a few 1960s experiments this application of AI is. Revenues from computer vision anti-fraud, facial recognition
to detect the edge of objects and categorize simple software, hardware and services are forecast to boom
shapes would spawn arguably the biggest trend in AI? from $1.1 billion in 2016 to $26.2 billion by 2025.1 This Health and life sciences: Predictive diagnosis,
means many of your competitors are either planning biomedical imaging, health monitoring
Those first forays into the field of computer vision to adopt computer vision as part of their AI strategy Manufacturing and energy: Supply chain
inspired modern neural networks that are now or are already working with it. optimization, automated defect detection
supporting an explosion of use cases in artificial
intelligence. You can find it in crime prevention and The availability of affordable high-performance Communications and retail: Customer
security – recognizing faces with incredible accuracy, compute power, the huge variety of visual data that intelligence, shopper measurement,
even from lampposts as suspects speed down roads pervades everyday life and more sophisticated fashion recommendations
in cars. Retailers can help you search for jeans worn by algorithms are driving computer vision uptake.
your favorite celebrity. In-store personalized advertising Think visual data is only generated in relatively small
and quicker checkout times are transforming the quantities? Forbes reports that 300 million photos This paper outlines what computer vision is and is not,
customer experience. Insurance firms can expedite are uploaded to Facebook every day and more than implementation considerations, and our thoughts
claims handling. Google will help you around foreign 95 million photos and videos are shared on Instagram.2 on how to achieve success based on our technology
locations by translating signs and other text captured and our work with pioneering customers.
by your smartphone. To make full use of this visual data – whether to
augment existing human processes, to radically change
the way your organization operates, or to come up with NEXT, WE TAKE A CLOSER LOOK
wholly new services, your teams, your lines of business AT WHAT COMPUTER VISION ACTUALLY IS
and your senior leaders must understand the use cases
of computer vision.
Share it with your lines of business
and executives
1 https://www.tractica.com/newsroom/press-releases/computer-vision-hardware-software-and-services-market-to-reach-26-2-billion-by-2025/
2 https://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#271d335760ba

2
In the spotlight
What is computer vision?
AT A GLANCE
Computer vision is an incredibly sophisticated tool.
The best way to appreciate this complexity is to
consider how human sight works. Human eyes have
In essence, computer vision is a subset
some six to seven million cone cells, each of which
of artificial intelligence that enables
contains one of three color-sensitive proteins, called
opsins. When light photons make contact with these computers to see and make sense of the
opsins a reaction occurs that generates electrical world. It is the AI application that allows a
signals. This message is transmitted to the brain computer to learn to analyze information
where it’s interpreted. from photos to video to thermal and
infrared data, amongst other sources,
Building an application capable of replicating that
and then make decisions or come to a
process is phenomenally intricate. Such a system would
have to be capable of processing and interpreting clear understanding of the environment
visual information so that it can be used for pattern and or situation based on that information.
object recognition, and for adapting 2D images from
our 3D world into 3D information. Yet in just a decade,
the accuracy of object identification rates has increased
from 50% to an incredible 99%.3 This makes computer NEXT, WE’LL REVIEW EXACTLY
vision more accurate than humans at reacting rapidly HOW COMPUTER VISION WORKS
to visual data.

3 https://www.sas.com/en_us/insights/analytics/computer-vision.html

3
Under the bonnet
How does computer vision work? Any training data will require within it sample images
of luggage containing prohibited items, such as
Computer vision is based on deep learning, a form
aerosols, weapons, and liquid containers, as well as
of machine learning that trains computers to perform
those containing permitted items. Each image will be
human-like tasks. Currently, these broadly include
tagged with metadata indicating the correct answer
understanding natural language, identifying and
– in this case, permitted or prohibited. A neural network
classifying objects in images or making predictions,
will process the visual data, using pattern recognition
but other uses are evolving. Rather than running
to identify the many different components of an image.
data through a series of predefined equations, deep
Its outputs, or ‘answers’ as to whether an item is allowed
learning establishes the structure of the model through
or not, are fed back into the system allowing it to
some basic parameters concerning the data, then
learn and improve in accuracy. So, instead of a human
trains the computer to learn on its own by recognizing
attributing certain characteristics to items, the machine
patterns using many layers of processing.
learns from the images it receives.

To emulate human sight, machines need to acquire,

process, analyze and understand images, which is
made possible due to the iterative learning process
of deep neural networks.

The process begins with a curated set of images or

video data, known as training data. This is used to help
the machines learn certain things about a certain topic; NEXT WE EXPLORE WHAT IS NOW
think damage to vehicles for the insurance industry POSSIBLE WITH COMPUTER VISION
or prohibited luggage items for the airline industry.

4
Beyond theory. AI can improve manufacturing 200,000 players analyzed 1,735,350 estimated new cases
detection rates by up to 90% for finding the next football of cancer diagnosed in the US
Into reality. Computer vision makes it possible star with AI in 2018
to spot defects not easily visible Computer vision makes it possible Computer vision helps identify areas
to the human eye to analyze every player, much to of concern in the livers and brains
enjoyment of the fans of cancer patients
Computer vision showcase
Airport security misses
Computer vision has more real-world
70% of unauthorized items $2 billion counterfeit bills 4 billion loss in US orange
applications than you might think – and
Computer vision enables safer in circulation in the USA alone market due to crop disease
the list is growing. The question many
CIOs and analytics leaders ask us is: will airline travel by identifying Computer vision makes it possible Computer vision makes it possible
the investment deliver groundbreaking unauthorized objects to spot counterfeit money and to detect early signs of plant disease
change? By that they mean faster and more prevent fraud to optimize crop yield
cost-effective ways to serve customers 93% decline in cheetah
and citizens, and innovative services that population over the past century 5-10mins is the maximum Proper oxygenation of the
differentiate them from competitors. Computer vision makes acceptable time customers placenta is essential for
Our answer is yes, but then we’ve been non-invasive monitoring of are prepared to wait in line a successful birth
working in the field of AI, specifically endangered species possible Computer vision makes automated Computer vision makes it possible
machine learning, for more than 40 years. checkout possible for a better to monitor oxygenation through
$460 billion counterfeit customer experience the umbilical cord
This collection of world vision projects
goods bought and sold
and their results highlights precisely how
around the world annually
transformational the technology can be.
Computer vision helps detect fake
goods and protects consumers

NOW, LET’S TAKE A LOOK AT THE STORIES

BEHIND THE HEADLINES…

5
Life-saving and Exponential possibilities An in-depth scan has a lot of data and

world-changing Diagnosing cancer, treating faster,

tracking outcomes
the outcomes can be improved with the
use of advanced analytics on patients’
health data and history.
In the Netherlands and USA, healthcare providers are
Computer vision in action actively embracing the potential of computer vision to GEERT KAZEMIER, CANCER CENTER AMSTERDAM

Any time a human could benefit from interpreting diagnose and expedite treatment recommendations.
lots of activities or a complex, or rapidly changing
Radiographers have a huge and growing workload, Phase 2. Augment
scenario, computer vision can help. As the
typically reviewing hundreds if not thousands of images
showcase illustrates, computer vision is especially Image data was combined with patient data such as
every day. Under that kind of time pressure and strain,
powerful when human eyesight may be fatigued their age, medical treatment, family history, and risk
it’s possible that suspicious structures or risk indicators
or unable to perceive the details of an image, factors such as diet and lifestyle. Our model reviews
could be missed. Computer vision is being deployed
especially where a large volume of visual data this visual and contextual information to determine
not to replace the experience of radiographers, it is
is being analyzed at speed. Insurance, policing, the probability that the scan shows cancer.
simply augmenting their efforts in order to accelerate
utilities, manufacturing are all able to ‘see things
review times and improve patient outcomes.
more clearly’ and make faster, more accurate Phase 3. Recommend
decisions because of computer vision. It offers We then built a recommendation engine that suggests
Let’s take a look at the process we developed for
us the potential to save lives and to change the treatments and interventions plotting them against risk
Amsterdam University Medical Centers (AUMC),
world as exemplified by these extraordinary and likely outcomes. For example, an elderly patient
whose physicians wanted a better, faster way to
stories. Use them as food for thought for your with multiple existing morbidities might be served
diagnose patients with liver and brain cancer. This
board, your stakeholders, your teams. better with a monitoring and palliative intervention
project demonstrates the transformational outcomes
that can be achieved, especially when computer vision than invasive surgery.
Please reach out to us to discuss how
is combined with other AI and analytical capabilities
we can help you build your computer
vision projects or develop your broader
such as NLP and recommendation engines. Phase 4. Learn
AI strategy. In order to ensure the system continues to perform at
Phase 1. Triage outstanding levels, the hospital then tracks the outcomes
With thousands of images to review, we built a model to of the recommendation engine, creating a feedback
REQUEST A CALL AT triage cases. The first step was to process images and loop that helps it to learn and further improve outputs
SAS.COM/COMPUTERVISION analyze constructs such as the boundaries of a tumor at both the triage and recommendation phases.
or lesions using a convolutional neural network (CNN).

6
Wild about conservation
Seeing patterns, predicting behavior, Conservationists at WildTrack knew that invasive
monitoring of black rhino populations in the 1990s had
protecting species
negatively impacted female fertility and they wanted
to avoid impacting cheetahs in the same way. They
Over the past 100 years, the wild cheetah population looked to computer vision as a way to remotely monitor
in Africa has declined by 93 percent. Just 7,100 remain cheetahs rather than performing the task manually,
with none now found in 76% of their former habitats. ‘on the ground’, as they had done previously.
Competition with humans for land has resulted in many Specifically, the team wanted to identify individuals,
being killed, and cubs are smuggled internationally their sex, age class, and their movements in order
– a journey that only one in six survives. to answer fundamental questions about what was
happening to their numbers, why, and what could
be done to prevent further population decline.

FIT for purpose

They created FIT, the Footprint Identification
Technique, that achieves the above ambition.
Images of footprints are analyzed by customized
mathematical models to provide the required
information. As WildTrack leverages crowdsourcing NEXT, WE’LL REVIEW SOME OF THE
to increase the volume of image data being gathered, CHALLENGES AND MISPERCEPTIONS
the progressive learning algorithms of deep learning THAT YOU MIGHT ENCOUNTER
that power computer vision can give the team a far WITH COMPUTER VISION
more accurate and in-depth view of what is happening
to cheetahs. Insights into behavior can augment and
scale the work being achieved by indigenous trackers Followed by practical advice on how to build
and be used to develop education programs and a cohesive approach to the development
practical conservation initiatives. and operation of this technology.

7
Challenges and misperceptions
Honesty is always the best policy • Computer vision does not come ‘out-of-the-box’ Sounds sensible enough, doesn’t it? In reality, however,
– models have to be trained before they can deliver many IT teams stumble when putting computer vision
Computer vision is clearly a powerful technology that
business value. Do you have the in-house skills or into production because they have not taken full
delivers value rapidly. However, as with any advanced
expert partnerships to bring computer vision use account of the infrastructure investment required
analytics program, there are misperceptions and
cases to life? If not, know which expert vendors can – from storage to compute and networking – to run
challenges that will require thought, planning, and
help you develop skills and roll out your projects. these kinds of advanced analytical functions.
investment from IT and data science leaders. Here is a
quick overview based on our experience with clients: • There’s no such thing as perfect model prediction From the data science perspective, data preparation and
– AI models are subject to decay like any other management alongside model management can often
• Computer vision is not a crystal ball
analytical model. Ongoing training will be required be overlooked or undertaken by different people using
to ensure the efficacy of this technology. a variety of tools. Not to mention the governance of the
• It is not going to take over decision-making
entire analytics lifecycle. Without proper governance
within your organization.
• Data preparation is crucial – many people believe and compliance processes in place, projects can
that data can be loaded into a model and answers quickly become highly complex, difficult to follow and
• It can’t create or rectify models on its own
will be generated. Incorrect. High quality, data impossible to audit. This not only has legal implications,
– because it doesn’t think on its own. This means
preparation is critical, nowhere more so than in it can also become very difficult for data science and
you cannot simply throw data into an AI system
training sets. Data must be rigorously cleansed and analytics teams to answer important questions, such as:
and have it build a model.
put into standardized formats – specialist activities Where is data coming from? Who is accessing it?
• Computer vision requires huge compute power that cannot be achieved automatically. It’s crucial What programming changes have been made and why?
to analyze visual data – what will this mean for your that you recruit data scientists with an appetite and
data infrastructure investments in future? At SAS, real interest in data management, and that your There is a solution that will give you the freedom
we collaborate with Intel and NVIDIA to support the data management strategy is governed by cohesive to be creative with this form of AI, while helping you
core elements of our AI offerings – partnerships that processes that everyone can adhere to. to maintain control of projects while in development
help to deliver faster, more accurate insights. and production.

NEXT, WE’LL EXPLORE THE SAS ANSWER

TO THESE CHALLENGES

8
See computer vision clearly
One platform for the entire This streamlined, governed capability comes in the
form of SAS Viya®, our unifying analytics platform that
analytics lifecycle
gives you the perfect balance of choice and control.
At SAS we are putting the automation and learning Importantly if offers:
of AI into the analytics lifecycle, where we begin with
data (acquisition and preparation), move to discovery • Support to load and process images
(experimenting with and building models), and finally or biomedical image data
deployment (where computer vision will be put into
your production environment). • Popular image filtration and color conversion
algorithms directly from SAS Viya

• Support for computer vision tasks using deep

learning, classification, object detection and
segmentation

• A friendly high-level Python package DLPy

available on GitHub

• Seamless Graphical Processing Unit (GPU)

support for model training and scoring

• SAS Event Stream Processing (ESP) for real-time

analysis of image or video data

9
We make all this possible
You need SAS Platform
with the SAS Platform (Viya) delivers
Run SAS Visual Data Mining and Machine Learning
(VDMML) on Viya and your organization will have a A full range of analytics, e.g.
high-performance solution that allows multiple users convolutional neural nets to
to explore vast sets of structured and unstructured autoencoder neural networks
data concurrently, via a simple visual interface.
Data engineers quickly build and run transformations, To explore data, move from
augment data and join data within the integrated visual simple models to highly
activity pipeline. All actions are performed in memory sophisticated ones
to maintain data structure consistency.
Governed processes,
You’ll access a broad suite of advanced statistical,
particularly if open source is
machine learning, deep learning and text analytics
used within your organization
algorithms, as well as analytical capabilities including
clustering, regressions, random forests, gradient
boosting models, support vector machines, natural To embrace open technologies
language processing, topic detection and others. and cloud with the ability
For expediency, you can test different modeling to move quickly to different
approaches together, compare results and put environments
the analytics into production across distributed Want to explore the possibilities
or traditional environments with auto-generated A streaming engine, for real- of computer vision yourself?
SAS code. time applications of AI, that
can be integrated into your
Finally, for maximum collaboration and flexibility, analytics solution so they can REQUEST A CALL AT
your people can access these capabilities from their talk to each other SAS.COM/COMPUTERVISION
preferred coding language.
To unify all technologies so
Here’s what our platform offers, mapped against
they are resilient and secure READ MORE ABOUT
your computer vision needs.
OUR CAPABILITIES
AND EXPERIENCE HERE

10
SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries.
® indicates USA registration. Other brand and product names are trademarks of their respective companies. Copyright © 2019, SAS Institute Inc. All rights reserved.