Preloader

OUR CLIENTS VALUE OUR

Projects

Machine Learning - CNNs

Objects detection

Target classification with FMCW radar signal

Enigma Pattern developed a method for training CNN models to classify objects in FMCW radar data. The team developed a unique way of processing domain-specific signals. The model manages to self classify versatile objects.

Technologies

Convolution Neural Networks, TensorFlow, scikit-learn, scikit-image

Results

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%

about us
Objects detection with FMCW radar signal
Machine Learning - CNNs

Detecting chemical compounds

Detecting chemical compounds in measurements of Ion Mobility Spectrometry

Enigma Pattern developed CNN models capable of detecting predetermined sets of chemical compounds in Ion Mobility Spectrometry data. The method is resilient to noise and changes in measurements due to different ambient conditions. At the same time it was important to determine which parts of the measured spectrum is the most important in the classification process.

Technologies

TensorFlow, Keras

Results

Enigma Pattern has identified a subset of features comprising 25% of the original information.
Prediction speed was increased 4 times.

Detecting chemical compounds
about us
Deep Learning

Objects classification

Deep Learning in classification of electronics items

One of the biggest distributors of Electronics in the World has employed Enigma Pattern to create a solution that will allow detection and classification of all 300,000 products using deep neural networks.

Challenges

The main challenge was the number of pictures available - 15 per SKU, the secondary challenge was related to the very high quality of those pictures as they were taken in a professional environment in perfect lighting conditions.

Technologies and methods used

Neural Networks, Keras, TensorFlow, Triplet Loss, VGG16, Computer Vision techniques

RESULTS

Classification of main categories: 97%
Classification of subcategories: 92%
Classification of a SKU: 87%

about us
Deep Learning in classification of electronics items
Deep Learning

detection and classification

Deep Learning in detection and classification of any LEGO elements

Enigma Pattern has created a new type of Deep Neural Network that allows real time detection and classification of LEGO items out of 400.000 different LEGO elements. For training it requires only synthetic images.

Challenges

  • Lack of real pictures of LEGO elements taken in different lighting conditions reflecting possible scenarios of children’s play.
  • Speed and accuracy of the selected method on mobile devices.

Technologies and methods used

TensorFlow, TensorFlow Lite, CoreML, Unity 3D, our proprietary “Synthesis” method

Results

mAP 89%, real time detection and classification experience

Deep Learning in detection and classification of any LEGO elements
about us
Supervised Machine Learning

Preventive maintenance

Preventive maintenance – No Trouble Found modeling

Using a continuous stream of information, individual to every mobile device, consisting of:
- internal state (OS, make, model, set of installed applications)
- user's behavior (number and time of phone calls, number and time of SMSs, WiFi on / off, etc.)

Enigma Pattern trained a model that predicts if a specific mobile device will crash in the near future.

Additionally the model was able to advise customer service on preventative maintenance actions such as upgrade/downgrade of OS, removal of applications etc. - internal state (OS, make, model, set of installed applications) - user's behavior (number and time of phone calls, number and time of SMSs, WiFi on / off, etc.)

Technologies and methods used

SHAP, decision trees (XGBoost)

Results

Model predicting failure of a mobile device with 91% accuracy.

about us
Preventive maintenance No Trouble Found modeling
Machine Learning - CNNs

Road Signs Detection

Road Signs Detection - Mapping systems for autonomous vehicles

Road mapping cars receive a vast quantity of visual data every second. This means that efficiency is key in the processing and analysis of the data.  Using neural networks our model enabled immediate visual recognition and segmentation of road-related signs and markings. These processes are essential for autonomous cars driving systems, as precise interpretation of road markings is critical to their successful operation. 

Technologies

Python, Keras, TensorFlow on GPU.
SSD Algorithms:  Single Shot MultiBox Detector

RESULTS
The system achieved 90% accuracy in visual signs recognition with set Jaccard Index parameters preserved.

Road Signs Detection
about us
Machine Learning - CNNs

Fixed Pattern Noise Removal

Fixed Pattern Noise Removal from Thermal Images

Thermal cameras are susceptible to both external (environmental) and internal (built-in) conditions. The objective of this project was to remove the fixed pattern noise.

Challenges

The primary concern was to remove the noise whilst preserving the real image. This meant that no additional data (hallucinations) should appear after noise removal. Since there are three different types of noises affecting the thermal images, each noise had to be removed separately.

Technologies

Keras, TensorFlow, Python

RESULTS
The low-frequency noise was decreased by 80%.
The number of artifacts was decreased by 30%.
The high-frequency noise was reduced by 20%.

about us
Fixed Pattern Noise Removal
Machine Learning - CNNs

Development of the process to train CNN by synthetic images

Development of the process to train CNN by synthetic images

Enigma Pattern was engaged to develop a unique method to generate and train convolutional neural network models basing on synthetic images. The overall process comprised:

  • development and preparation of the environment basing on Caffe
  • development of steering scripts responsible for simulation of the natural environment using Unity 3D
  • development of variants of the image object classifiers
  • modification of hyperparameters of the network to the improve detection precision
  • improvement of results by transformation of synthetic images 

Technologies

Caffe, Unity 3D, Python, Keras, TensorFlow

RESULTS
A repeatable process of building neural networks based on synthetic images

Development of the process to train CNN by synthetic images
about us
Supervised Machine Learning

Sound-based tires classification

Sound-based tires classification

Given only the audio recordings of spinning tires, the objective was to attribute each sample to a predefined class of tires (normal, under inflated, small-object-in-the-tire, etc.). Enigma Pattern developed a specially designed convolution neural network that learnt to properly distinguish between these different classes.

Technologies

Convolution neural networks (keras), specially designed FFT filters (numpy), Python

RESULTS
93% accuracy in tires classification.

about us
Sound based tires classification
CNNs with transfer learning and Hinton Capsules

Visual image processing

Visual image processing and images classification

Development of a neural network that would recognize benign tumors from set of CAT scan images.

The main challenge in recognizing images was the low signal data. We adopted an approach of transfer learning to pre-train the model and in the second stage the application of Hinton Capsules.

TECHNOLOGIES

Convolutional neural networks, with transfer learning and capsules network

RESULTS
CNNs – 71%
CNNs with transfer learning – 78%
Capsule Network – 84%

Visual image processing and images classification
about us
Polish National Health System

Heart Murmurs Detection

Heart Murmurs Detection

The main challenges was in recognizing sounds recorded by various types of stethoscopes.

A secondary consideration related to the wide range of a human’s heart beat depending on age, medical condition, and individual health profile. A neural network was developed to detect murmurs in heartbeats, which could be indicative of a problem, irrespective of the variable factors in recording or patient condition.

RESULTS
92% of recorded heart sounds were correctly validated as comprising heart murmurs or not.

about us
Heart murmurs