3D Deep Learning with TensorFlow 2
Computer Vision
Computer vision makes sight accessible to machines.
Automated image analysis
MethodologyOur sense of sight is considered our most important sense: About a quarter of the human brain is occupied with the perception and interpretation of light. Computer vision makes this sense accessible to machines as well. Unlike with humans, image data comes from a wide variety of sources, such as document scanners, smartphones, film cameras, high-performance industrial cameras or imaging processes from medical technology. Microwaves and infrared radiation as well as UV and X-rays are making things visible that remain hidden to the human eye. Depth images or point clouds can be analyzed to better map distances or volumes. In addition, there is a wide range of spatial resolution, from a few micrometers per pixel in microscopy to several meters per pixel in satellite images. This incredible diversity is matched by an equally impressive variety of processing methods ranging from classical image processing to Deep Learning. We help you to keep the oversight!
Use cases
Whether for industrial manufacturing, agriculture, medical technology, retail, remote sensing, automotive or in the media industry, computer vision can be used wherever visual analysis or decision-making is required. It often supports manual observations or automates routine processes to speed them up and enable objective as well as reproducible assessments.
However, computer vision models do not only aim at analyzing the actual state, but can also be used to predict events in the future: “Will the person cross the street?”, “When will the machine be defective?” or “How high will the harvest be?” are exemplary questions. In order to recognize changes as quickly as possible and to be able to react in time, videos are used instead of individual images.
The information extracted from images can be used in a variety of ways: Specialists react to defects found, rotten fruit is sorted out, robots bypass detected obstacles and doctors examine conspicuous tissue. Detected objects or image properties can also be further processed as metadata for recommender systems, language models can generate image descriptions using detected objects, and conversely, new images can also be generated using prompts.
How we work
Agile & transparentFrom the selection or creation of an image dataset, to the training of computer vision models, to the integration into the running operation – we consider the entire process chain together with you.
From the first proof of concept to the finished product with integrated MLOps pipelines, we work with you in an agile and transparent way to find the optimal solution.
Our experts in classical image processing as well as in Deep Learning select the appropriate models and architectures and train them in a targeted manner. We analyze the performance, investigate possible sources of error and optimize the solution approach until all requirements are also met during operation. Even limitations in the available hardware are no obstacle, as shown for example by our use case TinyML, where a model is optimized to get by with the memory capacity and computing power of a microcontroller.
You are not yet sure whether your problem can be solved with computer vision and which requirements have to be met? Then our Product Discovery Workshops are the right place for you.
Solutions from the field
spotsize: AI helps shoe shoppers determine the right size for a better user experience
spotsize has developed a solution which turns online shops into the shoe-shopping experience customers are familiar with from brick-and-mortar stores. Using an augmented reality (AR) solution, the spotsize application takes over the shoe-fitter’s role. With the help of machine learning, the program uses a smartphone camera to measure customers’ feet to determine the appropriate shoe size.
READ CASE STUDYThis is how computer vision can also be used at your company
Retail / Online sales
A large part of commerce takes place online, and the customer experience is enhanced by computer vision: product photos are classified, automatically sorted by category, and customers are suggested similar products based on images. In addition, product images can be generated or modified completely by AI. Realistic variants (other colors, sizes, living room background) are generated from a product photo of a sofa. In our case study with SpotSize we show how computer vision can be used to improve the online shopping experience: By scanning the customer’s feet at home and measuring them automatically via app, the optimal shoe size is determined and costly returns are avoided. Computer vision models can also be useful in offline business, for example to optimally align the sales floor with customer flows or for personalized purchase recommendations with smart mirrors.
Industry and manufacturing
In industry, computer vision can be used to check the quality of products, sort products or detect material defects – both in- and offline.
Production machines can also be continuously monitored using computer vision to detect whether early maintenance is necessary (predictive maintenance). Areas that are difficult to access, such as power lines, wind turbines or cranes, but also large industrial plants can be flown over with drones and the captured images analyzed to detect possible defects or perform remote maintenance. Robots in a production line can handle repetitive tasks as well as more complex ones if they are equipped with computer vision.
A typical challenge in industrial applications is the lack of training data suitable for the high-resolution cameras or special imaging techniques used. Generative AI helps here, e.g. by increasing the resolution with Super Resolution, by enriching data sets with generated images or by transferring them into a new domain, e.g. synthetic infrared images from RGB images.
You don’t find yourself in any of the industries? Feel free to contact us, whether you have an initial idea or already have a specific solution in mind – our experts will support you.
Florian Wilhelm
Head of Data Science, Contact for Data Management & Analytics
Hello! 👋
Get in touch!
Get in touch!
