So, what is computer vision? At the top level, computer vision uses computers, cameras and deep learning systems that work together to identify what they see and make, or help make, decisions based on what they see.
Decision making can be automated, but in most cases, consists of an alert system requesting human verification. Machine intelligence is the crucial component. This relies on trained algorithms that react to images scanned, empowering data-driven decisions in numerous situations.
How well does this perform? A recent test of Alibaba’s algorithm recorded an 81.26% accuracy rate in answering questions related to images, compared to a rate of 80.83% for humans.
How is the technology used?
While public awareness of these systems understandably focuses on person tracking through CCTV systems and semi-autonomous vehicles, computer vision has numerous real applications that already extend far beyond dystopian sci-fi scenarios.
In agriculture, satellite and drone-captured video and AI are used to help manage soil conditions, livestock and hydration management. Retailers use computer vision to manage warehouse inventory, checkout lines and stock control. Healthcare, defense, road traffic management...every industry is already making use of computer vision.
To view this video content, you must accept YouTube cookies.
These cookies make it possible to share or react directly on the social networks to which you are connected or to integrate content initially posted on these social networks. They also allow social networks to use your visits to our sites and applications for personalization and targeting advertising.
These many different use cases prove the deep value that computer vision may unlock for business. This alone makes it clear that concern regarding the use of these tools for surveillance should be tempered by the tangible benefits the technology may unlock. IDG describes these as process optimization (45%), product development (45%), augmentation of existing processes (41%) and the elimination of tedious or dangerous work (33%).
“By adding computer vision to components into organizational applications, enterprises can add significant value by increasing efficiency, automating business processes and/or reducing costs,” explains Dave Schubmehl, Research Director, AI Software Platforms at IDC.
Training takes time and data
Computer vision is not merely a process of pointing a camera at something and hoping for the best. Algorithms, such as convolutional neural networks, must be trained in a painstaking process using vast data sets to identify what they see correctly; they must then learn the correct decisions to take based on what they see. Many companies don’t yet understand the importance of training to build such solutions.
The need for data scientists to build these algorithms is driving demand beyond supply. QuantHub estimates that there was a shortage of 250,000 such employees in 2020. Salaries are high, and as enterprise customers recognize the potential benefits, demand continues to increase.
An important role for service providers
This has prompted the emergence of computer vision service providers, including Orange, which works with clients to build business-focused solutions or niche specialists such as Cromai or SWIR Vision Systems that serve the agricultural sector.
These providers help clients identify and develop solutions, including training computer vision using object classification or object detection techniques. The process exposes deep learning systems to huge collections of relevant images until they can reliably make appropriate decisions in response to them. Real-world tests must follow training and evaluation before systems go live, adding to project costs.
Given access to the skills required to build in-house solutions is limited, businesses seeking to explore computer vision will seek external support and must focus on precisely what business objectives they wish to achieve.
“Companies should start with setting the organizational groundwork, get the technology right, and capture and protect the value generated while scaling, and remain mindful of the context-specific factors and peculiarities that will influence successful AI adoption,” says Deloitte’s Navigating the Path of AI Adoption report.
At Schneider Electric, that core knowledge begins with training the machine-learning system to identify common faults in component manufacture alongside recognition of perfect units. For a retailer, knowledge may revolve around traffic flow. In each case, the AI you build will require access to large relevant data sets. Where can you source training images? How do you define project KPI? What problems do you hope to solve? As ever, the more focused your business can be at the start of the project, the more likely it will tend toward success.
This pre-development analysis may also expose opportunities to gather training images, such as placing cameras on production lines to train fault-reporting systems. Such collection may impact statutory privacy rights, which is why projects should be tuned to existing data strategy. Many enterprises will also invest in additional training data sets.
Once you have your system in place, it becomes necessary to iterate. Information your deep-learning system is trained to comprehend from what it sees may become less relevant over time. Once an enterprise builds an effective computer vision system, it must be prepared to maintain investment in the system to ensure that it remains functional.
These systems are not static. They reflect the challenges and opportunities of a changing world. Sunil Kumar Verma, Lead ICT analyst at GlobalData, says: “Increasing demand for vision-guided robotic systems to maintain production capacity and reduce dependence on the human workforce will further drive the adoption of machine vision usage areas.”