Statistical representation of wavelet transforms from low-feature ultrashort laser pulse spectrograms for improved neural network recognition.

摘要

The electric field associated with spectrograms generated by Frequency-Resolved Optical Gating (FROG) of ultrashort laser pulses can be recovered through an iterative computational process. The process, however, is limited in application by its long compute time. Training a neural network to recognize features in the spectrograms, or FROG traces, gives a more direct, or instantaneous, solution of the electric fields.;This thesis is a study of an original method of compact FROG trace feature description for neural network training. The method consists of performing a wavelet transform on each trace, and then describing groups of meaningful wavelet coefficients in each wavelet order through statistical moments in three dimensions. Experimental results demonstrate that this approach of using a wavelet transform as a basis for training a neural network on large low-feature FROG images is quite successful in terms of standard recognition error estimates.