Three parameters of Zwicker's loudness model, the excitation at threshold in quiet, the excitation that corresponds to the test signal and the reference intensity are researched in order to extract the loudness feature of signals. The sound intensity level at threshold in quiet state is introduced, corresponding to the centre frequency of the 24 critical-bands in equal-loudness contours. Then, an improved algorithm based on Zwicker's loudness model is put forward, with which the specific loudness values can be calculated quantitatively by using the energy of signals. Then a 24-D loudness characteristic vector is composed of the specific loudness values of the critical-bands, which can be used for signal analysis, mode classification and so on. The mode classification experiment of underwater targets is performed. The result shows that the improved algorithm based on Zwicker's loudness model is effective and practicable.