Copyright © 2021
To optimize your investments in new technologies, it is essential to know how to differentiate between traditional industrial vision and deep learning, and to understand their complementarity. This article sheds light on this area: how artificial intelligence can help assembly robots to identify the right parts, to detect whether a part is present, missing or incorrectly installed, etc.
Technological progress has multiplied over the past ten years: device mobility, Big data, artificial intelligence (AI), Internet of things, robotics, blockchain, 3D printing, machine vision, etc. The adoption of new technologies in the environment industry often requires their adaptation to specific constraints.
Let’s take a look at AI, and more particularly the analysis of images via deep learning using examples. Coupled with rules-based machine vision, it can help assembly robots identify the right parts, detect whether a part is present, missing, or incorrectly installed on a product, and determine more quickly whether the situation is a problem or not. In addition, all of these operations can be performed with great precision.
What is deep learning?
Like a neural network, the machine learns!
Take the example of graphics processors, called GPUs. A GPU brings together thousands of relatively simple processing cores, into a single chip, similar to that of a neural network. It enables the deployment of biologically inspired multi-level networks that mimic the human brain. It is therefore one of the forms of AI (Artificial Intelligence).
By relying on this type of architecture, deep learning makes it possible to carry out specific tasks without being expressly programmed for them. Where traditional computer applications are programmed by humans to perform given tasks, deep learning takes advantage of data (images, speech, text, numbers, etc.) to learn via neural networks. Starting from a primary logic developed during initial learning, neural networks continually improve their performance as they receive new data.
Sensitive to differences, never tiring
The process is based on detecting differences: it is constantly looking for changes and irregularities in a dataset. It is sensitive to unpredictable faults (a natural ability in humans), unlike computer systems that rely on rigid programming. Conversely, unlike a human inspecting a production line, the computer does not become weary of a repetitive task.
In everyday life, deep learning is more and more present: face recognition, recommendation engines on merchant sites, filtering of unwanted emails in electronic messaging, medical diagnostics, detection of frauds à la carte banking, etc.
In the process of being implemented in the industrial world
This technology is being incorporated as part of advanced production practices, including quality inspections and other use cases requiring decision making. Used to good effect in the factory in conjunction with machine vision, deep learning has tremendous potential for increased profits, especially when compared to investments in other emerging technologies, which may take years to become profitable.
How is deep learning complementary to industrial vision?
Machine vision: sensors based on fixed rules
An industrial vision system is based on a digital sensor integrated into an industrial camera equipped with specific machine vision lens. It allows you to acquire images transferred to a computer. Via specialized software, we can then process, analyze and measure various characteristics necessary for decision-making. Machine vision systems provide reliable results for regular parts, manufactured to consistent quality. They work using rule-based algorithms, filtering the different steps.
On a production line, a rule-based machine vision system can inspect hundreds or even thousands of parts per minute with great precision. It is more economical than inspection performed by humans. The conclusions drawn from the visual data are obtained through an automatic and rule-based method for solving inspection problems.
In the factory, classic rule-based machine vision is ideal for guidance (position, orientation), identification (bar codes, Datamatrix codes, markings, characters), measurement (comparison of distances according to given values) and inspection (faults and other problems such as the absence of a safety ring, a broken part, etc.).
Deep learning brings intelligence to machine vision
Rule-based machine vision works very effective with an established set of variables. Some examples: is a part present or absent? How far exactly is one object from another? Where should the robot collect this coin? These tasks are easy to perform on the assembly line in a controlled environment. However, in more nuanced situations, this technology is less suitable, which is where deep learning comes into play, thanks to the following advantages:
- Provides an answer to vision applications that are too complex to be solved using rule-based algorithms alone.
- Not disturbed by deceptive backgrounds or variations in the appearance of rooms.
- Allows applications to evolve by improving learning using new image data in the factory.
- Adapts to new examples without changing the structure of the programs.
A classic industrial example: looking for scratches on the screen of a device. This type of defect varies in size, location, and background type. By taking these variations into account, deep learning makes it possible to distinguish compliant products from defective ones. In addition, the assimilation of a new target (for example a new type of screen) can be carried out simply with the aid of reference images.
Inspection of visually similar parts with complex surface texture and varying appearance presents considerable difficulties for conventional rule-based machine vision systems. While functional defects almost always result in rejection, this is not always the case with cosmetic anomalies, which depend on the needs and preferences of the manufacturer. In addition, the latter are difficult to distinguish for conventional machine vision systems.
The advantages of deep learning for the industry
In summary, machine vision systems provide reliable results for regular parts, manufactured to repeatable quality, and applications become difficult to program as the number of exceptions and types of defects increases. For complex situations that require a vision close to that of humans as well as the speed and reliability of a computer, deep learning shows tremendous promise.
At the time of adopting new generation automation tools, traditional machine vision and image analysis via deep learning are therefore proving to be complementary, and not in competition. In certain applications, particularly those of measurement, rule-based industrial vision remains the preferred solution, because it is more economical. On the other hand, for complex inspections of products with many variations and unpredictable defects that prove impossible to program and perform using a traditional system, tools based on deep learning are a good alternative.
Read more: https://www.dzoptics.com/en/