MInD customized solutions
Outsorce your machine-learning work
Businesses in the industrial, medical, and consumer service sectors use our customized solutions to automate the processes that, until recently, only humans were able to do. The examples below illustrate what we can do with deep learning.
Industrial production lines need accurate, robust, and fast visual inspection solutions to find manufacturing errors. Real errors, such as cracks and surface inhomogeneities, are infrequent and varied, making the collection of sufficient training samples a challenge.
Training with synthetic data
We use the CAD model of the investigated product in a virtual reality environment and synthetically generate all possible types of errors. The MInD solution trained on photorealistic synthetic images makes a decision on a real production line with 99.5% accuracy in 100 milliseconds.
Oncologists assess the negative side effects of radiation therapy by classifying chromosomes in the metaphase. First, an operator looks for 10 different metaphases on a sample at 100x magnification. For each metaphase, they then switch to 1000x magnification to count all the chromosomes and look for various aberrations. Evaluating a sample takes about 15 minutes.
The MInD solution counts chromosomes with a greater than 99.9% accuracy and alerts the oncologist to potentially abnormal chromosomes. Evaluating a sample takes less than a minute, on average.
MIRA Measure offers professional opticians phone and tablet apps to customize and optimize eyeglass lenses. Opticians need to find the exact boundary between the lens and the frame on an image and locate the pupil relative to the identified contour.
Finding the contour of lenses
The cloud-based MInD solution localizes the pupil with 99.5% accuracy, even on blurred photos. It finds the contour of a lens in about 30 milliseconds with 99% accuracy, even under weak lighting conditions.
Gaming advisors analyze long hours of gameplay videos of their clients to collect accurate statistics on their strengths and weaknesses.
For example, they have to read the graphical “kill feed,” in which real-time information is provided about who killed whom with what kind of weapon.
The MInD solution can read the “kill feed” twice as fast as and 10 times more accurately than classic computer vision algorithms.
Real-time motion analysis software is an important tool for professional trainers. Most systems use markers to measure the time evolution of joint angles and require a special set of cameras to track markers.
Measuring joint angles
Using a single camera, the markerless motion analysis engine tracks joint angles with high accuracy in real time.
Vegetables and fruits have different shapes and colors that often make it hard to spot abnormalities.
Sellers want to pick out damaged products early to prevent the degradation of other items.
The MInD anomaly detection model alerts sellers to potential problems by showing the location, size, and extent of an anomaly.
Estimating the number of people in potentially crowded spaces is important to keep them comfortable and safe and avoid overcrowding.
Estimating a crowd
The MI-CROWD component of the MInD platform allows for a real-time estimation of the number of people in a space with 95% accuracy, so traffic can be redirected to less crowded spaces in a timely manner.
The nutrient and water content of arable soil is finite, meaning that crops and weeds can only grow in competition with each other.
Accordingly, both the average yield and the quality of the crops greatly depend on the timing and success of weed control.
The MInD solution can identify 12 different weeds with 90% accuracy, even in the seedling phase. By identifying the distribution of weeds on arable land at an early stage of development, farmers can protect their crops using targeted, thus fewer, chemicals.
Coated tablets may have surface defects that are hard to detect through human visual inspection.
Checking surface quality
The MInD platform offers tools to identify anomalies that the human eye cannot see, including color defects, shape deviations, and surface roughness.
RDefects in high-voltage transmission lines, including cracked insulators or corroded joints, pose a threat to humans and the environment. The human visual inspection of the quality of the components is cumbersome.
Analyzing drone images
A deep learning-based object detection model can provide input for moving a drone closer to the components. If a drone has a clear shot, the automatic defect-analyzing system provides a more accurate evaluation than a human operator.
In physiotherapy, immediate feedback from a therapist is necessary for a patient to avoid re-injury and stay motivated. As a consequence of the COVID-19 pandemic, access to physical rehabilitation services has become limited.
With our deep learning–based program, we can measure the time series of the patient’s joint angles on the camera image. The patient also has the option of wearing an accelerometer and a gyroscope when exercising. Both the patient and the physiotherapist can receive immediate feedback from the image and time series data on whether the patient is performing the exercises as directed.