Machine vision and quality control

What the machine vision service includes

We develop the software part of a visual quality-control system and set it up on the customer's equipment. For the specific task we collect a dataset of defects, label it, train a recognition model and integrate it with the client's cameras, lighting and line. The camera, conveyor and actuators themselves remain the customer's property and physically perform the imaging and rejection, while we prepare the algorithm that makes the "pass or fail" decision on a frame. In the end you get a trained model tied to your hardware, plus the reaction logic: mark a part, stop the belt, log the defect. The work also includes calibration for the shop floor — angles, shadows, flow speed — so machine vision delivers a stable result not on test shots but on real products.

How visual control really works

A frame from the camera goes through preprocessing (brightness normalization, alignment, noise filtering) and is then fed into a neural network that detects objects and classifies their state. The detector finds zones of interest in the image — a seam, a hole, a marking, a surface — and rates each by trained features, outputting coordinates and a defect probability. The result is compared against an allowable threshold, and the system forms a signal that the customer's equipment executes: it sorts, marks or stops. Speed is critical here — modern detectors make an inference in a single pass over the network, so control runs at the pace of the line without slowing it down. Importantly, the model works not by hard rules like "if a pixel is darker, it's a defect" but by the statistics of thousands of examples, so it catches defects that are hard to describe with a formula.

Where computer vision came from

The field grew out of academic work of the 1960s: in 1963 Lawrence Roberts defended a dissertation at MIT on reconstructing the three-dimensional shape of objects from a flat image, which is considered one of the starting points. In 1966 Marvin Minsky and Seymour Papert launched the educational Summer Vision Project at MIT — by the memo dated 7 July, the task of separating objects from the background was set as a summer assignment for students, and its underestimated difficulty set the tone of research for decades. The breakthrough in convolutional networks came in 1998, when Yann LeCun and co-authors published the paper "Gradient-Based Learning Applied to Document Recognition" and the LeNet-5 architecture for recognizing handwritten digits. The mass leap was driven by the AlexNet network in 2012 (Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton), which cut the top-5 error on the ImageNet contest to about 16 percent versus 26 for the nearest competitor. In June 2015 Joseph Redmon and colleagues presented the YOLO algorithm (arXiv paper 1506.02640), which recognizes objects in a single pass over the network and made detection suitable for real time.

Why software and precision decide everything

In a quality-control system the physical equipment is the same for everyone — what sets the result apart is precisely the trained model and its setup. If the dataset is collected crookedly or the threshold is set at random, machine vision will either let defects through or reject good parts, and both scenarios hit the economics of production. Accuracy is measured not by an overall percentage of "correct answers" but by the balance between missed defects and false triggers, and this balance is tuned separately for each task. The model has to be retrained for specific materials, lighting and camera wear, otherwise on a new batch of products the accuracy drifts. That's why the main engineering work lies not in buying hardware but in labeling, training, validating and calibrating the software part for your shop floor.

What tools we work with

The base stack is OpenCV for preprocessing and frame geometry, neural-network detectors of the YOLO family for real-time object search, and deep-learning frameworks for training and retraining the models. OpenCV covers the classic operations: camera calibration, distortion correction, contour extraction, work with color spaces — what's needed before the frame goes into the network. YOLO and related architectures are responsible for detecting and classifying defects with an allowance for the line speed. We train the neural networks on the customer's labeled dataset, and we deploy inference where it's convenient by latency and cost — on an industrial PC at the line or on a server. The specific set is picked for the task: somewhere the classic OpenCV methods are enough, somewhere there's no doing without a trainable neural network.

When the key tools of the stack appeared

The OpenCV library was started at Intel in 1999 under the direction of Gary Bradski, and the first public release came out in 2000 — since then it's been the de facto standard for computer-vision tasks. The Python language, on which most of the work on training models and tying the components together is done today, was released by Guido van Rossum in 1991. Convolutional networks as a practical tool were established by the LeNet-5 architecture in 1998, while the turning point toward deep learning was set by AlexNet in 2012. The YOLO detector, on which real-time recognition is built, was published in June 2015. These dates show that the tools aren't raw: OpenCV and Python have been working for over two decades, and neural-network detection has been battle-tested since the mid-2010s.

Why you can trust this to us

The combined experience of our development team exceeds 45 years in IT, and our approach is engineering rather than demonstrative: the model is verified on a held-out sample and a test batch before it goes into the production launch on the line. We don't sell hardware and don't manufacture your products — we write the program and set up the equipment you already have, so we're responsible precisely for the part where accuracy is decided. Before launch we run the model on real frames from your shop floor, record the percentage of misses and false triggers and agree the thresholds with you rather than setting them by eye. If on a new batch or a change of lighting the recognition quality drops, the model is retrained — this is a standard procedure built into the maintenance. As a result you get machine vision tied to your production and verified on your data, not a universal off-the-shelf solution.