Studying the performances of linear signature detector on a time-frequency representation

This paper deals with automatic recognition of time-frequency signatures. In particular, we are interested in linear structures in the time-frequency plane. These signatures are detected by their boundary in the time-frequency plane. Then, shapes are compared to a rectangular model using a distance evaluation on Fourier descriptors. The value of the distance allows the decision if the considered boundary is a linear signature or not. The performance of this method are evaluated on "linear chirps" by tracing the received operating characteristic (ROC) curves. Finally, our method is applied to the recognition of real "hydraulic noises" in underwater acoutics.