Ambient Noise Imaging; Enhanced Spatial Correlation Algorithms and a way
to Combine Independent Images of Improved Stability and False-Alarm Rejection
A0079Choon Kiat Lim Acoustic Research Laboratory, Tropical Marine Science Institute, NUS
A0080John R. Potter Acoustic Research Laboratory, Tropical Marine Science Institute, NUS
In the growing class of underwater Ambient Noise Imaging (ANI)
algorithms, three stochastic techniques have so far been proposed and
. The original first-order Acoustic Daylight algorithm (which images
the mean received intensity as a function of look direction)
. Second-order temporal intensity imaging (which uses the standard
deviation of the intensity)
. Second-order spatial intensity imaging (which calculates spatial
cross correlations of pixel intensity at zero time lags).
This paper addresses the challenges of second-order spatial imaging,
and how to combine images formed by independent algorithms. The
original second-order spatial algorithm calculates the correlation
matrix of beams for some fixed time window and takes the matrix
element with the smallest coefficient to identify two seed channels.
The extent to which other beams correlate to these two seed channels
is then used to form the image. Two enhancements to this base
technique are proposed. Furthermore, we propose a method of combining
images from different algorithms based on Kmean clustering.
The first image algorithm enhancement is based on the original spatial
correlation algorithm, except that the procedure is repeated with the
seed channels reselected from a small group of leading candidates each
time. For each selection, an image is produced. The final image is
constructed as a weighted product of these images. Results indicate
that this algorithm produces considerable improvement (as much as 15dB
in some cases).
The second enhancement attempts to classify the channels into target
and non-target sets in such a way as to maximise the intra-correlation
within each set while simultaneously minimising the inter-correlation
between the sets. The seed channels, corresponding to the matrix
element with the smallest cross correlation, are found first. The
beams are then classified into target and non-target sets by either a
simple clustering method or setting a correlation coefficient
threshold. The intra and inter-correlations are then modelled as a
linear objective optimisation with non-linear and linear constraints.
The aim is to find a representative time series signal for each set by
maximising the intra-correlation within each set. Finally, the time
series of the seed channels are replaced by their respective
representative time series and the image is generated by finding the
new correlation matrix. The procedure is repeated until convergence.
Results indicate only slight improvements in image quality when
compared to the original method (2 dB or less).
Lastly, this paper proposes a method of combining images produced by
different algorithms (i.e. Acoustic Daylight, second-order temporal
and spatial imaging). We apply Kmean clustering, with a validity
measure. The result obtained by a particular algorithm is assigned to
an element of a feature vector. The best results are obtained by
first normalizing the input feature for each algorithm to zero mean
and unit variance. The elements within each feature vector are then
sorted before clustering. The final image is obtained by a weighted
sum of the individual images, the weights being determined by the
clustering algorithm. Results show an improvement of up to 4dB
compared to the clustering method without sorting.