The use of non-supervised neural networks to detect lines in lofargram
J.-C. Di Martino, Brigitte Colnet, Marcello Di Martino · 2002
The topic of the article falls into the pattern recognition issue. More particularly, it deals with the extraction of spectral lines from a lofargram of low signal to noise ratio using unsupervised neural networks. The approach is based on constrained Kohonen's self-organising maps able to include the perceptual relevant features of spectral lines. This approach is well suited to enhance lines whatever their shapes and without using any information about signal features (stationarity, duration, number of narrow-band components) and noise level. Experimental results concerning a set of lofargrams at different signal to noise ratio prove that the approach has a good robustness to noise.>