Skip to main content

Research Theme: Signal Processing and AI

Research Theme: Signal Processing and AI

Advanced Intelligent Signal Processing for the next generation of autonomous systems.

Our research in Advanced Intelligent Signal Processing is world renowned. We provide cutting edge, explainable, interpretable and verifiable Artificial Intelligence (AI) for the next generation of autonomous systems. Our research covers the development of novel theory and applications in the fields of public security and healthcare.

Our world leading research informs our teaching, supervision and mentoring. We produce highly employable, multi-skilled graduates and postgraduates. Our graduates are very successful and working all over the globe. We supervise and teach talented students to the highest standards through our PhD, MSc, MEng and BEng Programmes.

Signal and Information Processing: digital fingerprint in neon blue surrounded by 1s and 0s and a circuit board

We are actively involved and regularly contribute to the world premier forums in Signal Processing. We have international and national collaborations within academia and industry. Our research has led to technology transfer and exploitations in security and surveillance, biometrics, and medical technologies. Our research projects have resulted in spin-out companies in the creative industries.

We have received funding from research councils and industry, including:

  • Engineering and Physical Sciences Research Council (EPSRC)
  • Biotechnology and Biological Sciences Research Council (BBSRC)
  • Medical Research Council (MRC)
  • Wellcome Trust
  • EU FP7 (7th Framework Programme)
  • InnovateUK
  • Newton Fund
  • Fera Science
  • Ministry of Defence (MoD)
  • Thales
  • Neura Technologies
  • Neura Health Group Ltd

Intelligent Health

We have a strong tradition of developing image analysis techniques. The techniques that we develop have potential for high impact in clinical practice. We take concepts from feasibility studies through to clinical translation and commercial exploitation. We work with clinical and industrial partners.

Our research uses advances in AI and machine learning. It has led to the development of new systems that can revolutionise healthcare. These include new approaches to:

  • understanding health risks
  • predicting disease progression
  • creating personalised health interventions for improved patient outcomes

These approaches allow us to develop innovative tools to support clinicians and reduce spending.

We can achieve this tremendous potential by developing effective, trustworthy machine learning applications. These applications should be implementable in real healthcare contexts.

Intelligent Health: diabetic retinopathy

We work with multi-disciplinary groups of machine learning experts, clinicians and engineers. Together, we develop human-centred machine learning solutions that transform care pathways and improve health outcomes.

Our current projects on Intelligent Health include:

  • Next Generation 3D histology for earlier prostate cancer detection
  • Advanced applications of ocular optical coherence tomography (VELOCITY)
  • Diabetic retinopathy detection using machine learning
  • Investigation into mitochondrial dysfunction in facial appearance and ageing
  • Automated detection and assessment of pain and distress in preterm babies using deep learning
  • Application of artificial intelligence to predict effectiveness and adverse events to biologic and non-biological systemic therapy in psoriasis
  • Fuzzy neural inference system applied for the classification of cancerous colonic mucosa cell images
  • Prediction of nodal spread of breast cancer by using an artificial neural network

Machine Learning for Signal Processing

Our research focuses on the mathematical foundations and practical applications of signal processing algorithms. The algorithms are able to learn, reason and act. It bridges the boundary between theory and application. It allows for development of novel theoretically-inspired methodologies. These target both longstanding and emergent signal processing applications.

The core of MLSP lies in its use of intelligent, nonlinear and non-Gaussian signal processing methodologies. It combines these with convex and non-convex optimisation.

Machine Learning: heartbeat - EEG

MLSP encompasses new theoretical frameworks for statistical signal processing. These include:

  • neural networks
  • Hidden Markov Models (HMMs)
  • independent component analysis (ICA)
  • nonnegative matrix factorisation (NMF)
  • Tensor factorisation
  • Bayesian methods

MLSP couples these with information theoretic learning.

Novel developments in these areas are specialised to the processing of a variety of signal modalities. This includes audio, speech, bio-signals, images, multispectral, and video.

More than statistics

Statistics is organised information, but intelligence is more than that. It is not enough to solve a problem by giving the most ‘frequent’ answer. We are developing theories and demonstrable algorithms for statistical machine learning that truly integrate signal processing and computational intelligence.

Our algorithms are able to discover knowledge for themselves. They learn from new information whenever unseen data is captured. Many researchers discount the need for human intervention in data analysis. Our work goes further. It provides the freedom to combine human intuition and machine intelligence.

Our current projects in Machine Learning for Signal Processing include:

  • signal separation (eg single- and multi-channel recordings, audio source separation, bio-signal separation including EEG, ECG and MEG)
  • speech recognition, activity recognition, event detection
  • inference and prediction of hidden patterns in signals/images
  • compressive sensing and sparse information retrieval
  • biomedical data analysis
  • computational social scene analysis
  • pervasive audio-visual sensing

Computer Vision and Biometrics Recognition

Our research takes place at the cutting-edge of security and intelligent surveillance technologies.

Predictions acknowledge that by the year 2020, global databases will hold information on over a billion people. This will result in many challenges to the whole subject area. These challenges mainly relate to the reliability of the systems. This is particularly true when the raw image data contains imperfections. These can include poor light, imperfect equipment, or even deliberate and fraudulent alteration.

Computer Vision and Biometric Recognition: eye

Computer vision

Our Computer Vision research leads the international field in the extraction of object behaviour models and dynamic face models from image sequences and real time video. Our work has been widely applied to people and object tracking, human gesture and pattern recognition.

Our current projects in computer vision include:

  • statistical machine learning
  • time series analysis
  • multi-modal data fusion
  • biologically inspired vision
  • image compression
  • multi-camera networks
  • image analysis
  • video tracking
  • 3D reconstruction from 2D images
  • automated monitoring of lameness in dairy cattle

Biometric recognition

The work involves multi-modal techniques and fusion of data at the feature level. Our models can be applied to biometric recognition systems. They have led to efficiencies in complexity, speed and accuracy. These efficiencies result in recognition rates which show a robustness against variation in image size. This is due to efficient image feature extraction and classification.

Current projects in biometrics recognition include:

  • sclera recognition
  • iris recognition
  • face recognition
  • fingerprint
  • palmprint
  • voice recognition
  • keystroke
  • signature
  • body movement
  • gait biometrics

Multimodal Information Processing

Signal and Information Processing is particularly well suited to deal with multimodal data. Multimodal data includes audio, video, and infrared. Current and future sensor systems, such as the sensors in smart devices, will provide ever more data for analysis.

The next generation of artificially intelligent systems include:

  • automated security and surveillance
  • human anomaly detection
  • human identification and tracking in cluttered and congested environments
  • speech enhancement and separation in challenging environments

Systems such as these will need to process multimodal data.

We develop algorithms and technologies to enable significant advances towards this vision.

Multimodal Data Processing - human detection

Multimodal human anomaly detection

Human anomaly detection is a challenging task for reliable artificial intelligence.

Human behaviour, both as individuals and in groups, is complex. Such behaviour needs robust features for modelling, indexing, and classification.

Our research into human behaviour analysis includes:

  • multimodal action recognition
  • contextual information retrieval
  • information fusion

We are investigating advanced deep learning techniques. Our main focus is to provide network explainability and human-informed systems.

Multimodal data processing - human behaviour anomaly

Multimodal human identification and tracking

Our research involves:

  • core signal processing for multiple human localisation
  • tracking in cluttered and congested environments

Core signal processing techniques provide us with the solution for Multimodal Human Tracking (MHT). These techniques include:

  • social interaction modelling
  • probabilistic clustering
  • non-linear filtering
  • efficient data association
  • particle flow based on stochastic differential equations
  • multi-level cooperative fusion
Multimodal Data Processing - human tracking

Multimodal speech processing

We have established complete mathematical proofs for nonlinear signal processing theory. These are major breakthroughs. From this, we have established non-linear models for optimal signal separation and information retrieval.

A major success is our development of a complete framework using statistical methodologies for:

  • signal fusion
  • separation of complex nonlinearly mixed signals
  • informational retrieval

This work has radically challenged conventional approaches. Research for speech and audio show outstanding performance for various cases:

  • over-determined, where the number of channels is larger than the number of sources
  • determined, where both are equal
  • under-determined, where the number of channels is smaller than the number of sources

Under-determined cases include binaural and single channels.

The work has contributed to the fundamentals of nonlinear signal processing theory. It has exploited advanced machine learning techniques. It paves the way for the research community to adopt and develop interdisciplinary solutions.