Faster, cheaper automated image analysis
A technique for automated image analysis based on our research is used today to inspect industrial components, draw data from medical images and track and analyse faces.
CyberOptic's annual revenues for automated optical inspection systems.
Most conventional methods of image analysis are expensive to develop and require bespoke algorithms for specific applications.
However, researchers from the Centre for Imaging Science and School of Computer Science pioneered a radically new approach that quickly 'trains' the analytical system without having to tell it what to look for. The system automatically builds a 'model' against which it can compare similar images from the same genre.
Specific applications of our image analysis techniques have generated multi-million dollar revenues for several spin-off companies, which have commercialised specific applications of this technology: Kestra for automated inspection of printed circuit boards (PCBS), imorphics for medical image analysis, and Genemation for face image analysis. Staff from the School independently established two additional spin-offs: Image Metrics and Optasia.
Inspecting printed circuits boards
Techniques that place components directly onto circuit boards, known as surface mounting, markedly reduce the costs and dimensions of digital electronics. However, component mounting has a high error rate, so it is important that every circuit board is inspected for mistakes.
Automated optical inspection (AOI) of circuited boards based on our research proved to be more accurate and simpler than previous methods. Kestra developed a commercial system which experienced rapid sales, leading to Kestra's acquisition by CyberOptics. By 2011-12, CyberOptics' annual revenues from the Kestra AOI system had reached $14.3m.
A similar system was also developed for number plate recognition. Although this is a mature market, the superior performance of our image analysis has supported sales of approximately 100 units worth $1.5 million.
Analysis of medical images
We developed techniques to interpret sophisticated 2D and 3D medical images using a new way of identifying the boundaries between tissue surfaces. Our algorithms can analyse output from all major medical imaging systems (Siemens, Philips and GE). Siemens, for example, uses our algorithms in its 3D/4D cardiac ultrasound system. The company has a 12% share of this $1.25 billion market.
Faceware Technology revenue
Faceware Technology has an annual revenue of $7 million.
The medical applications of our image analysis research are commercialised by spin-out company imorphics. Software from this company is being used to develop surgical tools and brain scanning techniques. It also provides image analysis technology for the drug trials of four global pharmaceutical companies. Imorphics has an annual turnover of £1 million.
Similar techniques are also used in a commercial system to calculate the age of bone from X-ray images. This package is used today in 22 major children's hospitals across northern Europe and is responsible for around 13,500 diagnoses each year.
Microsoft Kinect for Windows (KfW), available in 39 countries, incorporates our work on active shape models (ASM) and active appearance models (AAM). A KfW development kit, which has been downloaded 500,000 times, allows software developers to access face tracking and analysis for their applications.
Two postgraduate students from our School established Image Metrics (now Faceware Technology) which captures facial motion in films and video games. The technology has been used in major films (including The Wolfman, Meet Dave, and The Curious Case of Benjamin Button), the computer game Grand Theft Auto IV which grossed $500 million in its first week, and Red Dead Redemption, which sold more than five million copies within its first two weeks. Faceware Technology has an annual turnover of around $7 million.
Researchers at the School of Computer Science developed a new generic approach to automated image interpretation. Researchers found they could use statistical techniques to quickly and cheaply develop specific image analysis software by training a generic system with sets of similar images from the area of application.
Key research findings:
- Statistical models of the shapes and spatial arrangements of structures in a given class of images can be learned and used to interpret unseen images automatically.
- Statistical modelling can model the shape and photo-realistic appearance of structures within a class of images.
- Shape models can be built automatically, without any manual annotation of 'landmarks' in training images.
- These fully automated methods can create powerful image analysis functions for a wide range of applications including industrial inspection, medical image analysis and face tracking.