About This Research Area

The same efficient-AI methods solve problems well outside the clinic — on farms, on highways, and in documents.

Livestock monitoring and welfare

About the Project

System diagram with a video decoder, pig detector, tracker and head pose classifier.
The detection and tracking pipeline.

On-farm welfare assessment is subjective, slow and prone to bias. We automate it.

  • One system photographs pig carcasses from dorsal and lateral views and scores welfare indicators
  • Tail length, a key indicator, is measured by segmenting the tail and fitting a polynomial curve, converting pixels to centimetres using the animal's body length
  • A second system estimates pig pose and analyses behaviour from barn cameras

Ferri et al., Computers and Electronics in Agriculture, 2024.

Diagram of a pig with a tail bounding box and the steps used to segment and measure tail length.

Tail length estimation for carcass-based welfare assessment.

Overhead camera view of pigs in a barn pen, used as input for pose estimation.

Pig pose estimation and behaviour analysis.

Crop yield prediction from hyperspectral imagery

About the Project

Workflow from drone data acquisition through a deep learning model to deployment on an edge device.
From drone capture to an edge device in the field.

Predicting yield early helps farmers plan. We predict canola yield from drone-captured hyperspectral images.

  • A 1D CNN across 150 bands and eight dates reaches R² = 0.82
  • Feature selection, pruning and quantization shrink it until it runs on an edge device in the field
  • A related project predicts yield in aeroponic systems and uses SHAP analysis to show which growing conditions matter most

Valarezo-Plaza et al., IEEE Transactions on AgriFood Electronics, 2024.

Horizontal bar chart ranking growing conditions such as crop type and humidity by importance.

Which growing conditions most affect predicted aeroponic yield.

Licence plates and commercial vehicles

About the Project

Table comparing detection accuracy and processing time for three licence plate localization algorithms.
Localization results on the NTUA Medialab database.

Reading a licence plate is trivial for a person and hard for a machine, especially at highway speed on embedded hardware.

  • Plates localised and recognised with a MobileNet-style network using depthwise separable convolutions
  • Accelerated with OpenVINO on a Raspberry Pi 3
  • A three-year project extends this to commercial vehicle inspection: HAZMAT placards (85.0% mAP over 15 classes) and CVSA decals at highway check stops

Super-resolution

About the Project

Network diagram upscaling a small face image through residual and upsampling blocks to a large one.
Reconstructing a 128×128 face from a 16×16 input.

MSG-CapsGAN was the first capsule GAN applied to super-resolution, reconstructing a 128×128 face from a 16×16 input and beating the state of the art on PSNR and SSIM. The same idea extends to airborne magnetic maps, where high-resolution data is scarce.

Majdabadi and Ko, Multimedia Tools and Applications, 2020.

Document and text understanding

About the Project

A text passage beside extracted entity-relation triples and generated descriptive sentences.
Extracting descriptive relations from text.

Scientific literature is growing faster than anyone can read it.

  • We detect and extract structural regions from academic papers
  • SciBERT+ and SciBERTN perform named-entity recognition over scientific text
  • We created the Descriptive Relation Dataset (DReD): 3,283 annotated paragraphs describing relations between noun phrases

Also in this area

Further work

  • Synthetic-data training for deformable linear object segmentation, teaching robots to see wires
  • Fast detection of explosives on embedded hardware
  • Erosion detection in pipelines
  • Autonomous robotics, including leader-follower formation control and the UV-SAFE disinfection device