Eight-Year Work Summary of a Machine Vision R&D Engineer

I've worked in the machine vision industry for 8 years, specializing in research and development. In this summary, I’ll discuss some technical features of the industry, offering insights and guidance for newcomers.

What is Machine Vision?

Machine vision isn’t actually about "vision" as humans understand it. It lacks human visual comprehension. Instead, it's an engineering application of image processing technology. Tasks are carried out by algorithms and optical hardware developed by engineers. Each algorithm is designed to complete a specific task, with little to no cross-application.

Common Misconceptions about Machine Vision

Some believe that machine intelligence has surpassed human intelligence, but that's overly optimistic. Current technology is still far from this level. Machine vision can only solve structured, simple inspection tasks with reference mathematical models, like shape matching, edge detection, and texture recognition.

If there's no discernible pattern to detect, the vision system can't be developed. While humans can quickly identify a particular object, even in complex backgrounds, machines struggle with this. For instance, finding a specific item like a button or glove within intricate patterns can be challenging for machines because engineers can't find reliable features to create an effective system.

The Advantages of Machine Vision

However, machine vision does have its strengths. It excels at repetitive tasks, like detecting scratches on glass or defects on screens, with high accuracy and precision.

Machine vision essentially uses machines to replace human eyes in measurement and judgment. Applications typically involve system integration or secondary development, categorized into four main areas:

1. Appearance and Defect Inspection

   • Relies on template matching to detect appearance issues and defects.

2. Recognition

   • Involves biometric recognition (e.g., face, voice, fingerprint, iris), target recognition (e.g., license plate, radio frequency), barcode recognition (1D, 2D), character recognition, and texture recognition. Recognition ultimately aims to classify, requiring big data training and deep learning.

3. Dimensional Measurement

   • Measures geometric dimensions (e.g., length, width, height, circumference, area, volume), and shapes like circles or ellipses. Calibration is crucial, involving camera calibration.

4. Positioning

   • Used for workpiece alignment, assembly, robotic picking, and stacking. Machine vision systems quickly acquire large amounts of information, allowing for automatic processing and integration with design and manufacturing control data. This makes them valuable for monitoring operations, product inspection, and quality control in modern automated production.

Thoughts on Artificial Intelligence

Current AI in machine vision merely discovers methods for automatic pattern extraction. It collects enough representative samples, learns from them, then uses the system to classify images. It isn't truly "intelligent"; it's just a technology, still within the realm of Turing machines. The gap between AI and genuine understanding of images is significant, providing plenty of research opportunities for scientists, master’s, and doctoral students.