In manufacturing, machine vision is primarily used in several key areas, including visual guidance, dimension measurement, product inspection, and object recognition. Among these applications, one of the most fundamental algorithms is product recognition and localization. For instance, in visual-guided robots, it’s essential to first identify the target product in an image and precisely determine its coordinates before guiding the robot to the correct position. The same principle applies to dimension measurement and product inspection: before performing any measurements or inspections, we must first confirm whether a product is present and where it is located, so that subsequent analytical tools can be properly applied. Therefore, product recognition and localization represent a fundamental challenge.

01

Components of a visual positioning system

The robot positioning system based on machine vision comprises a camera system and a control system. The camera system includes a computer (equipped with an image acquisition card) and a camera, which primarily collect visual images and apply machine vision algorithms. The control system consists of a control box and a computer, and it controls the precise position of the computer’s end-effector.

The workspace uses a CCD camera for image capture and employs a computer to recognize the images, extracting tracking features and performing data computation and identification. By leveraging inverse kinematics, the system determines the error at each robot position and then controls the high-precision end-effector module, scientifically adjusting the robot’s position and orientation.

02

Key Factors of Visual Positioning Systems

In the field of industrial production, especially in the application of industrial robots,Visual Recognition and Localization SystemThis is particularly important. In actual production, what we need to focus on is not only whether we can grasp objects accurately, but also the speed at which we do so. This has long been a persistent challenge in the industry—typically, the grasping speed of industrial robots we encounter is rather slow. Yet once we try to increase the speed, the precision of the grasp starts to suffer. This is precisely the crux of the challenge faced by visual recognition and positioning systems. Next, let’s join Xiao Ju as we explore this topic further.First is the data volume,In a relatively complex production environment, the system needs to accurately locate and identify the products that require recognition.Next is speed,How can we boost the speed of some standard production lines to the millisecond level? While previous algorithms could still function effectively under normal conditions, as algorithms continue to evolve, deep learning algorithms often require more powerful GPUs to achieve optimal performance.Then comes the heart of the issue: positioning accuracy.In deep learning systems, the images we see are all subject to some degree of scaling. Therefore, the entire system needs to achieve pixel-level alignment with the original image. What remains is the accuracy of recognition. In many cases, we have very limited labeled training data. Under these circumstances, how can we further improve the accuracy of recognition?

03

Visual localization faces challenges.

To design a feasible algorithm for product recognition and localization, several challenges need to be addressed:   1. Rapid identification of products—Industrial products vary widely and are highly diverse. Therefore, for each specific application, it’s essential to quickly identify the target product from just a few images—or even a single image. For example, if a production line needs to locate rivets, taking one photo and performing initial training allows the algorithm to subsequently search for and locate these rivets in subsequent images.   2. Fast product search—For a 2-million-pixel image, the algorithm typically needs to identify and localize the product’s position within tens of milliseconds.   3. High-precision localization—Industrial production has strict requirements for accuracy and tolerances; thus, product localization must strive for maximum precision. Today, it’s common to demand that recognition and localization algorithms achieve pixel-level precision, and sometimes even sub-pixel precision.   4. Adaptability to product defects—If a product is partially occluded or obscured, resulting in missing parts in the image, the algorithm still needs to accurately detect and locate the object. Conversely, if the product surface becomes dirty or contaminated, altering its surface features, the algorithm must remain capable of recognizing and locating the object despite these changes.   5. Robustness to uneven lighting—If the illumination on the product varies—for instance, half of the product is brightly lit while the other half is dim—the algorithm should still be able to reliably detect and localize the product.   6. Ability to recognize rotated products—Products often rotate freely within a 360-degree range.   7. Multi-product recognition—An image may contain multiple products simultaneously, and the algorithm must be able to identify and localize each one separately.   8. Accurate recognition of nearly symmetrical objects—Nearly symmetrical objects can easily be misidentified if not properly handled; therefore, the algorithm must incorporate appropriate design considerations to avoid such errors.   9. Handling object polarity reversals—For example, if the trained model assumes a white background with black text, but the actual product image might have a black background with white text, the algorithm must still be able to recognize and differentiate between these two cases.