Method Research On Auditory Perception Based Underwater Target Recognition

The automatic recognition of underwater targets is a crucial technique in the modern naval war,which is still an unsolved problem because of the complexity of oceanic background and the variability of targets.The remarkable abilities of sonar operators about discriminating sounding objects motivate the researchers to address this problem by mimicking the auditory perceptual principles.Based on such an idea,this dissertation conducts systematic studies on auditory feature extraction(including timbre features and central auditory representations),discrimination strategy analysis,and the applications of both aspects into target recognition.The main contents are as follows.1.The relationship between auditory perceptual principles and sound target recognition is explored.By summarizing the research progress of timbre spaces,auditory models,and decision making processes in human discriminating objects,the theoretical foundations of applying auditory perceptual principles into underwater target recognition are established.Combing with the special properties of underwater targets,the detailed methods are proposed.2.Underwater targets are divided into multi-levels of categories,and the timbre attributes of them are investigated in a hierarchal manner.Three groups of paired comparison experiments are designed including 1)man-made and natural targets;2)submarines,ships,and patrol boats;3)various types of ships.The CLASCAL model is used to analyze the obtained dissimilarity ratings and thus 3 timbre spaces are found corresponded to the three experiments.The first dimensions of the first two spaces separately distinguish man-made and natural targets as well as submarines,ships,and patrol boats,whereas the differences among various ships should be represented by a 6-dimensional space.Finally the correlation or regression analyses are used to extract the classification features in each layer and all of them are validated.3.The central auditory model proposed by Shamma is applied into both the timbre modeling of underwater noises and the discrimination between man-made and natural targets.Combining the central auditory representations with PLS and RPLS respectively,the tasks about timbre modeling and target recognition are completed.The results show that the central auditory model exhibits superiorities in performance compared to those of traditional timbre descriptors.4.The performances and strategies of human listeners and automatic classifiers are compared in various discrimination tasks.Three groups of subjective discrimination experiments including 1)man-made and natural targets;2)ships and submarines;and 3)three types of ships are conducted.The results indicate that the performances of automatic classifiers perform better than those of the human subjects in all tasks but the discrimination between man-made and natural targets,and the strategies employed by excellent human subjects are similar to that of logistic regression.Automatic classifiers and several human subjects demonstrate similar performances when discriminating man-made and natural targets but in this case,their strategies are not similar.An appropriate fusion of their strategies leads to further improvement in recognition accuracy.5.The robustness of central auditory representations and timbre attribute features are evaluated by a classification experiment.The experiment includes 9 usual target categories,and 4 different types of additional noises are used for interference.The deep neural networks are also combined with the two auditory features to obtain good performances.The results indicate that the central auditory representations have better discrimination power and robustness,which depend on the noise inhibition of the peripheral stage and the target/noise separability of the central stage.6.A feature transfer learning method based on neural networks is proposed.The invariant features can be automatically learned from the prototypes of underwater targets in the daily environment and then are used in underwater target classification.The results indicate that the learned features outperform the original ones and such superiorities are much obvious when the training set is small.