Deep Learning in Computer Vision
Deep Learning has made Computer Vision techniques viable in many more areas than before. There was already a good deal of industry knowledge and experience around detecting and classifying objects in the visible portion of the electromagnetic spectrum, but not so much in other areas.
Digica has developed expertise that opens up new domains like the InfraRed and near-InfraRed, Ion-Mobility spectrometry, micro-Doppler signals from FMCW radar, and the audible spectrum.
Objects detection with FMCW radar
What the customer wanted to achieve
FMCW radar is good at detecting intrusions into the airspace of busy airports, but not so good at differentiating between objects of a similar size like drones and birds. Since birds will always be there and drones should only be there by invitation, the customer needed to be able to tell the difference between them, fast. False alarms can trigger crippling disruption to operations, so they wanted a system that would be able to distinguish between birds and drones reliably.
How Digica helped the customer
The team developed a unique way of processing domain-specific signals in FMCW radar data to train convolutional neural networks (CNNs) to successfully classify objects.
What we achieved
Scalar data (azimuth, distance, speed, etc.), 1s observation
gain in Recall: 5%-10%
Scalar data + microDoppler spectra, 1s observation
gain in Recall: 20%-25%
Scalar data + microDoppler spectra, 5s observation
gain in Recall: 30%
Technologies used
Convolution Neural Networks, TensorFlow, SVM, scikit-learn, scikit-image
Detecting chemical compounds
What the customer wanted to achieve
Ion mobility spectrometry is used for detecting dangerous substances, but it only works when you can shield the data from the adverse effects of noise and different conditions. This usually takes a lot of analysis and a lot of time, both of which are costly, so the customer wanted quicker analysis results that were at least as accurate as traditional methods.
How Digica helped the customer
We developed CNN models that can detect predetermined sets of chemical compounds in ion mobility spectrometry data. Despite unwanted noise and measurement fluctuations that can occur in different ambient conditions, our method is still able to identify which parts of the measured spectrum are the most important for classification.
What we achieved
Digica identified a subset of features which amount to just 25% of the original information, so prediction speed was increased to 400% with no loss of accuracy.
Technologies used
TensorFlow, Keras, CNNs
Sound-based tires classification
What the customer wanted to achieve
A major tire manufacturer wanted an efficient way to detect when truck tyres need maintenance. The system would use gate-mounted microphones to capture the sound of passing vehicles.
How Digica helped the customer
Our objective was to take the captured audio samples of tyres on the move and match them with established classes of tires in different states, such as normal, under-inflated, small object embedded, and so on. Digica developed a custom convolution neural network that learnt how to accurately distinguish between tyres in these different states.
What we achieved
The system classifies tyres with 93% accuracy.
Technologies used
Convolution Neural Networks (Keras), specially designed FFT filters (numpy), Python
Heart murmurs detection
What the customer wanted to achieve
The customer wanted to overcome the challenges involved with recognising heartbeat characteristics. The sound of a heart changes when you listen with different kinds of stethoscopes and factors like the patient’s age and aspects of their general health can also modify the sound in unhelpful ways too.
How Digica helped the customer
We developed a neural network that can detect potentially problematic heart murmurs, despite all the factors that can obscure this information in the audio.
What we achieved
The system successfully identified heartbeat characteristics in 92% of recordings.