In these fields, one of the most fundamental algorithms is product recognition and localization. For example, in vision-guided robots, the robot needs to identify the product to be grasped in an image and pinpoint its coordinates before guiding it to the product's location. The same applies to dimensional measurement and product inspection; before measurement and inspection, it's essential to know whether a product exists and where it is located before applying subsequent analytical tools. Therefore, product recognition and localization is a fundamental problem.
01
A robot positioning system based on machine vision comprises a camera system and a control system. The camera system includes a computer (with an image acquisition card) and cameras, primarily collecting visual images and applying machine vision algorithms. The control system includes a control box and a computer, controlling the specific position of the end effector. The working area uses a CCD camera to capture images, and the computer recognizes the images to obtain tracking features, performing data calculation and recognition. Inverse kinematics is used to obtain the error at each position of the robot, and then the high-precision end effector module is controlled to scientifically adjust the robot's position and pose.
02、
In industrial production, especially in the application of industrial robots, visual recognition and positioning systems are of paramount importance. In actual production, we need to focus not only on accurate grasping but also on speed, a persistent challenge in the industry. Industrial robots often exhibit relatively slow grasping speeds. Increasing speed, however, compromising grasping accuracy. This is a major hurdle for visual recognition and positioning systems. Let's explore this together. First, there's the issue of data volume. In complex production environments, the system needs to accurately locate the products requiring recognition and positioning. Second, there's the speed issue. How can we achieve millisecond-level speeds on standard production lines? While older algorithms might work, deep learning algorithms often require more sophisticated GPUs for implementation. Then comes the core problem: positioning accuracy. In deep learning systems, images are scaled, requiring the system to achieve pixel-level matching with the original image. Finally, there's the issue of recognition accuracy. In many cases, the available training data is limited. How can we further improve recognition accuracy under these circumstances?
04、
To design a feasible product identification and location algorithm, several difficulties need to be overcome:
1. Quick Product Identification: Industrial products vary greatly. Therefore, for each specific application, it is necessary to quickly identify the product to be searched from several images, or even just one. For example, if the current production line needs to locate the position of a rivet, a photo can be taken and the relevant data can be learned, allowing for searching and location in subsequent images.
2. Fast product search: For a 2-megapixel image, it is usually required to be able to identify and locate the product within tens of milliseconds.
3. High-precision positioning: Industrial production has strict requirements for precision and tolerance, so product positioning must be as accurate as possible. Currently, it is generally required that positioning algorithms can achieve a positioning accuracy of one pixel, or even sub-pixel level.
4. Adaptable to the effects of missing, occluded, or dirty products: If a product is occluded, causing a certain proportion of it to be missing from the image, it should still be able to be identified and located. Conversely, if the product surface is dirty, causing changes in surface features, it should still be able to be identified and located.
5. It can adapt to the effects of uneven lighting. If the brightness of the product changes, such as half being bright and half being dark, it still needs to be able to identify and locate.
6. Products that can be rotated: Products can typically rotate within a 360-degree range.
7. Can identify multiple products. There may be multiple products in an image, which need to be identified and located separately.
8. It can accurately identify nearly symmetrical objects. Nearly symmetrical objects are easily misidentified as being in the wrong orientation, requiring appropriate design adjustments.
9. It can handle polarity reversal of objects. For example, the product being learned is white with black text, but the actual product image may be black with white text, which needs to be recognized.
