Locating algorithms for the implementation of the initial windowing algorithms was trivial. The majority of these are well documented and are easily implemented. Dolph-Chebyshev window algorithm, however, was a little more elusive. Initial attempts to implement this algorithm were unsuccessful. Two different approaches were taken; initially an attempt was made to implement this purely in the time domain, based on an algorithm sourced from the speech recognition research at the Institute for Signal and Information Processing, ISIP, located at Mississippi State University. After many hours spent trying to validate the algorithm, it was decided that this could not be used. A second attempt was made using a window generated in the frequency domain, using code based on the original 1947 paper by Dolph. This algorithm, however, only supported odd window sizes, and the implementation of the IFFT library contained in the GDMS required that its window size be 2n and thus this algorithm had to be discarded. The final algorithm that was implemented is based on FORTRAN code published in 1977 by IEEE Acoustics, Speech, and Signal Processing Society, The windowing subsystem is designed to be extensible. Using OO techniques, a general algorithm for generating the window is combined with a specialised class which generates each individual coefficient. This means that to add another algorithm costs as little 10 lines of code, including user interface integration.

eqimgwindowing.png

For all windows, coefficients can be generated for an arbitrary window size. The requirement that the windowed data be transformed to the Fourier domain imposes further limitation in that the Fourier libraries we are using will only operate on a data set of size 2n, 0 < n < ∞. This restriction has been implemented at the User interface level, maintaining the flexibility of the windowing subsystem. Due to the simple nature of the majority of the algorithms in this module, speed is not generally an issue. With the exception of Chebyshev, the most expensive calculation involves two cosine and four multiplication operations. Since the size of the windows used tend to be small this is a relatively insignificant overhead. Dolph-Chebyshev is a little different, as this window is calculated in the Fourier domain, then transformed into the time domain via an IFFT it is computationally more expensive. In order to support the flexibility of the windowing subsystem, this algorithm implements its own IFFT. Due to the small size of the window used, however, this cost is minimal. In fact, it is more than offset by the increase in flexibility and confidence in the integrity of the algorithm. To address performance concerns and to speed up the process as a whole, the number of calculations performed in this subsystem has been carefully considered and recalculation of the coefficients is only performed when it is unavoidable. The set of algorithms supplied in this release is by no means exhaustive. There are some notable absences, such as the Kaiser window, which would be a valuable addition to the GDMS. Due to the flexible design of this subsystem, the addition of another algorithm can be achieved at minimal expense and is encouraged. Another avenue which might be explored with regard to enhancements would be the simplification and optimisation of the Dolph-Chebyshev algorithm. This was a port of a FORTRAN function and was implemented faithfully as maintenance of data integrity the primary concern. This algorithm uses an inverse Fourier transform internally and for windows of significant size this could prove to be expensive. Optimisation of this might prove to be beneficial if window sizes are becoming large.