Vision Systems and Deep Learning: Are We Headed for Summer or Winter?

Spring is the time of new beginnings. Plants sprout and blossom. Summer is the season of growth and maturity that leads to the harvest. In the earth’s climate, summer always follows spring. In winter, plants wither and go dormant, animals hibernate, and the world becomes still and cold. Activity diminishes.

In business and technology, there is no such order of seasons. Winter can follow spring just as easily as summer.

Both artificial intelligence (AI) and machine vision have experienced economic winters. AI has had two winters: the first from 1974 through 1980 and the second from 1987 through 1993. Machine vision experienced its own winter in the mid-1980s. During the winters, funding for research and startup capital dried up. Customers, disappointed by unfulfilled expectations, became disinterested in adopting the technology.

We can trace the causes of these winters in both AI and machine vision to closely related causes:

• New technology that promised to open up new markets.
• Rapid growth of entrepreneurs (or researchers) looking to capitalize on the opportunity.
• Flow of investment funds into the new enterprises and research endeavors.
• Promises made by entrepreneurs and researchers to attract the funding.
• Customers clamoring to adopt the new technology and reap its benefits.
• Unfulfilled promises to investors, research funding sources, and customers.

We stand on the threshold of a perfect storm where all of these characteristics are present again.

Nothing has impacted the world of vision systems over the last decade as much as the emergence of deep learning (DL). DL is latest big advance in artificial intelligence (AI). Not only has DL made some difficult problems easier, but it also has allowed vision system technology to extend into new areas that were not tractable with legacy image analysis. The lure of programming with data—that is, images—rather than writing code is seductive. Promises of easier programming and the entry into new fields attracted many entrepreneurs and lots of investment capital and spawned numerous startups with hardware and software to make DL. It also excited many potential customers.

We have brilliant people developing DL solutions, but who also often have limited experience in real-world challenges. We have startup companies that raised significant capital based on business models that make promises to investors and the market that may not be practical. And, we have integrators and end users who are anxious to move along with the latest technology whenever possible—even before it’s proven as commercially viable.

We might be facing an economic and technical winter for DL. Fortunately, though, if we learn from the past, the winter is not inevitable. Here are some actions that will help prevent it.

Prioritize High-Quality Images

We have the technology to acquire high-quality images in most situations. These are images that have the right resolution, high contrast or high dynamic range, and low noise. DL is seductive in its ability to show promise with low-quality images. With enough training after acquiring a very large set of images, the results from DL may almost be acceptable. However, the cost of acquiring and labeling an extensive set of training images and conducting many training sessions will far outweigh the cost of ensuring a high-quality image. Quality images require fewer training images and training cycles to achieve superior performance over what can be achieved with low-quality images.

As everyone learns in Computer Science 101A: garbage in, garbage out.

Don’t Misapply DL

DL seems to be the solution to a wide range of applications. There are often other less complex and less costly solutions. Implementing a DL solution for an application that could be solved more easily and with less cost leaves the DL solution open to valid criticism.

There is a wealth of sensing technologies that don’t require imaging and work extremely well. A good system integrator is aware of all options. Applications that can be solved with straightforward coding using existing software packages are almost always a simpler, less costly, and more reliable solutions than DL.

Misapplication of DL leads to it being relegated to the solution of last resort or maybe the solution of no resort.

A DL Solution Takes a Team

The DL solution team needs to include people with excellent expertise in applying DL. That’s not enough though. It also needs people who understand the industrial process, procedures, and environment. Team members knowledgeable in industrial practices can spot missing information in a requirements document and identify corner cases where the vision system must operate sensibly. A DL solution not only needs to work well in the lab—it needs to satisfy the customer in the field and build trust with the customer.

Be Responsible for the Whole Solution

When a project fails to live up to expectations, every party involved shares all the blame. Although a team of DL experts might join up with an experienced systems integrator, if the solution fails to work or works well but not well enough, the end customer will expect all members of the team to contribute to the needed improvements.

That means the DL practitioners need to understand what the integrator is doing and why. The integrator needs to understand what is being performed in the development of the DL portion and why. Not only does this avoid problems, but it also provides the opportunity for synergy.