Home About Snape Signals Signal
Processing Pattern Recognition Data Mining Research
Papers Mindfulness
·
Enhancement of ultrasound images of stainless steel
components
Ultrasound is routinely used to check
for the presence of defects in welded steel components. B-scan and C-scan
images are built up from a raster of A-scans. Defects are seen as regions of
high backscatter in the image. This approach is effective for ferritic steels
but can be difficult in stainless steel because the metal grain size is
comparable with defect size. Moreover, austenitic steel grain structure is not
homogenous. Snape Signals has developed linear and non-linear image filtering
methods that make defects stand out from random grain scatter. The most
successful methods are based around Gaussian decomposition, rank order, and
pixel connectivity filters. For more information see slides: Enhancement
of ultrasound images (303kB), Ultrasound
Papers (14kB)
·
Adaptive Non-linear Filters
Signal and image processing systems
often use adaptive linear filters for system identification or inverse
filtering. Practical examples are echo cancellation and channel equalisation.
Many systems are non-linear and ideally require an adaptive non-linear filter
for system identification or inverse modelling. The ubiquitous neural net – the
MLP – is sometimes used in this role, but learning is generally long and
un-reliable. An alternative robust non-linear filter is based on n-tuple
sampling and a single layer “perceptron-like” structure. The system, called the Single Layer Look-Up
Perceptron, has strong connections to the WISARD system developed by Igor Aleksander et al for face recognition. It has been used in
applications as diverse as non-linear Wiener filtering, to formant extraction
for automatic speaker recognition. For more information about Single Layer
Look-Up Perceptrons see: Single Layer Look- Up Perceptron
(202kB), Non Linear Filtering Papers (21kB)
·
Noise Robust In-Car Speech Recognition
Speech
recognition systems have many potential applications for secondary control in
cars. Robust recognition in the high noise environment of a car is still
difficult. One of the sources of noise
is in-car entertainment. Unlike other noise sources, the recognition system can
have direct electrical access to this noise as it is generated. Snape Signals has worked with a speech
recognition company to develop algorithms that make the recognition process
less sensitive to in-car entertainment noise. The method is based on spectral
subtraction within the pattern-matching core of the recogniser. For more detail
see: In-car
noise cancellation paper (148kB), Speech Recognition Papers (26kB)
Home About Snape Signals Signal
Processing Pattern Recognition Data Mining Research
Papers