Skip to main content

Artificial Intelligence

Computer Vision

The progress vision-based technologies achieved by the use of Artificial Intelligence is nothing short of a miracle. We can now detect fires, recognise people by their faces and spot cancer on tissue samples. With speed and precision beyond human capabilities.

What we do

Newest Computer Vision methods

At Digica, we use the newest Computer Vision methods, in conjunction with Artificial Intelligence to any data and problems that can be translated into images or movies. We detect different objects (people, vehicles, drones, etc.) both in visible light, in thermal, and near-infrared spectra.

We work with radar data, where we use radio frequencies electromagnetic fields to build images of landscapes and weather, but also to detect different objects from far away or from a very close distance as we do in gesture recognition projects.

We use 2d and 3d imaging for chemical projects, where we analyze signals from spectrometers and other chemical devices.

Case studies

 
Development of the process to train CNN by synthetic images

Development of the process to train CNN by synthetic images

 

What the customer wanted to achieve

The customer wanted a system that can detect objects like vehicles, people, and animals in different situations such as in the street, the sky, or underwater. We couldn’t use the usual tens or even hundreds of thousands of pictures to train the system, which was an extra challenge.

How Digica helped the customer

We developed a unique method of generating and training convolutional neural network models based on synthetic images. For the overall process:

  • we developed and prepared the environment, basing it on the Caffe deep learning framework
  • developed steering scripts to simulate the natural environment using Unity 3D
  • developed variants of the image object classifiers
  • modified network hyperparameters to improve detection precision
  • transformed synthetic images to improve results

What we achieved

Created a repeatable process for building neural networks based on synthetic images.

Technologies used

Caffe, Unity 3D, Python, Keras, TensorFlow

 

Detecting people in masks on thermal images during COVID crisis

Detecting people in masks on thermal images during COVID crisis

 

What the customer wanted to achieve

Facial recognition technology has come a long way, but during the COVID-19 crisis systems with infrared cameras were failing to detect many people.

The customer wanted to help during the COVID-19 pandemic with a temperature scanning station that brought together infrared cameras with facial recognition technology. The tricky part was that surgical masks being worn to limit the spread of infection were confusing to existing systems.

How Digica helped the customer

We used RGB pictures of people in masks and thermal images that we created in-house using an adaptation technique. Once the detector had been trained using these image sets it was able to identify people wearing masks and send their location to the temperature measuring algorithm. We could describe the result as an ‘automated remote thermometer’ and it helped scanning stations to prevent the spread of COVID-19.

What we achieved

Recall increased from 90% to over 99%.

Technologies used

TensorFlow, MobileNet, SSD

 

Road signs detection

Road signs detection

 

What the customer wanted to achieve

The customer wanted accurate road sign detection for autonomous vehicles that would still run on devices with low computing power.

How Digica helped the customer

Road-mapping cars receive a vast quantity of visual data every second. This means that efficiency is key in the processing and analysis of that data. Using neural networks our model was able to immediately recognise and segment road signs and markings. These processes are essential for autonomous car’s driving systems, as the precise interpretation of road markings is critical to their successful operation. 

What we achieved

The system recognised road signs with 90% accuracy while preserving set Jaccard Index parameters.

Technologies used

Python, Keras, TensorFlow on GPU.

SSD Algorithms:  Single Shot MultiBox Detector

 

Fixed pattern noise removal

Fixed pattern noise removal

 

What the customer wanted to achieve

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.

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

How Digica helped the customer

Digica developed a method which was customized and adjusted for the very specific conditions that apply to the images collected by the customer.

What we achieved

Decreased low-frequency noise by 80%.

Decreased the number of artefacts by 30%.

Reduced high-frequency noise by 20%.

Technologies used

Keras, TensorFlow, Python

Visual image processing - CNNs with transfer learning and Hinton capsules

Visual image processing - CNNs with transfer learning and Hinton capsules

 

What the customer wanted to achieve

The customer wanted to be able to recognise benign tumours from a set of CAT scan images. The main challenge was the low signal data.

How Digica helped the customer

Digica developed a neural network to recognise benign tumours on a set of CAT scans.

We adopted a transfer learning approach to pre-train the model and apply Hinton capsules in the second stage.

What we achieved

CNNs – 71%

CNNs with transfer learning – 78%

Capsule Network – 84%

Technologies used

Convolutional neural networks, with transfer learning and capsules network

How can we help you?

 
To find out more about Digica, or to discuss how we may be of service to you, please get in touch.