Efficient Visual Intelligence Technologies

The project aims at developing computationally efficient methods, algorithms and tools in the area of intelligent image processing. Examples include but are not limited to efficient real-time object detection, recognition, and false positives suppression in live video streams. In this project, we focus on dealing with computational and theoretical issues imposed by inherent high-dimensionality of data, scarceness of samples, data uncertainty, and robustness to groups of transformations.

The project started in 2011 as a joint research and development venture between Apical LTD (now a part of ARM) and Department of Mathematics. It has attracted funding from

  • Technology Strategy Board (Knowledge Transfer Partnership agreement KTP009890 between the University of Leicester and Apical Limited, now part of ARM)
  • Technology Strategy Board (Knowledge Transfer Partnership agreement KTP010522 between the University of Leicester and Visual Management Systems Ltd)
  • European Regional Development Fund (Partnership in Knowledge Transfer between the University of Leicester and Apical Limited, now part of ARM)
  • Direct support from Apical Limited in terms of hardware, sponsorship of PhD students, and co-supervision of MSc students.

Machine Learning Algorithms in High Dimensions

As a part of this project we  studied the problem of efficient ``on the fly'' tuning of existing, or legacy, Artificial Intelligence (AI) systems. The legacy AI systems are allowed to be of arbitrary class.  They may include Support Vector Machines, Deep Learning Neural Networks, Decision Trees, or any other meaningful tool or their combination. They must, however, share the following property:  the data they are using for computing interim or final decision responses should posses an underlying structure of a high-dimensional topological real vector space. We showed that dealing with occasional errors in these systems can easily be achieved without complete re-training. Instead of re-training a simple cascade of perceptron nodes is added to the legacy system. The added cascade modulates the AI legacy system's decisions. If applied repeatedly, the process results in a network of modulating rules ``dressing up'' and improving performance of existing AI systems. Mathematical rationale behind the method is based on the fundamental property of measure concentration in high dimensional spaces. Further details can be found in:

A.N. Gorban, R. Burton, I. Romanenko, I.Y. Tyukin. One-trial Correction of Legacy AI Systems and Stochastic Separation Theorems.  2016.  https://arxiv.org/abs/1610.00494

Example 1. One-trial learning

A brief demo showing capabilities of the current system (Apical/ARM) to learn and detect new objects on-the-fly (using our specially developed Scanner App) is provided below

At the University of Leicester, the project is carried out at the Visual Intelligence Laboratory (VisLab) and involves several members of staff (Dr. I. Tyukin, Prof. A.N. Gorban, Dr. E. Mirkes, and K. Leschke), Research Students (S. Green, K. Sofeikov), and a KTP Research Associate (R. Burton, ARM co-supervisor is Dr. I. Romanenko)

Example 2. One-trial removal of false positives

Our test video (NOTTINGHAM) comprising of 435 frames was taken from the streets of Nottingham using an action camera.

 

When passed through a pre-trained classifier built using Deep Learning Convolutional Neural Network (VGG-11A), it generates 189 false positives.

In order to see these false positives we created the following video. Objects that are detected correctly are marked as magenta ellipsoids. False positives are shown as yellow rectangles.

 

To deal with these false positives we used our one-trial learning algorithm.  The algorithm uses one dot product (corrector). Filtered and non-filtered videos, stacked on top of each other for comparison, are provided below. Performance of the corrected system is shown in the top video. Performance of the original system with false positives marked by yellow rectangles is shown in the second video (right below the first one).

As we can see, single perceptron is capable of removing false positives returned by standard deep learning convolutional neural network from "one-shoot" or one-trial learning. Performance of a one-trial learning single-element corrector is further illustrated with the figure below. The figure shows the number of false positives removed vs the number of true positives that have been accidentally removed too. The data is provided for the NOTTINHGAM video above.

 fp_rate

Share this page:

Contact details

Department of Mathematics
University of Leicester
University Road
Leicester LE1 7RH
United Kingdom

Tel.: +44 (0)116 252 3917
Fax: +44 (0)116 252 3915

Campus Based Courses

Undergraduate: mathsug@le.ac.uk
Postgraduate Taught: mathspg@le.ac.uk

Postgraduate Research: pgrmaths@le.ac.uk

Distance Learning Course  

Actuarial Science:

sep-dl@le.ac.uk  

 

DisabledGo logo

The University of Leicester is committed to equal access to our facilities. DisabledGo has detailed accessibility guides for College House and the Michael Atiyah Building.