선박 분류를 위한 합성곱 신경망의 전처리 성능에 관한 연구

Abstract

The problem of vessel classification has been actively studied for vessel design, underwater target detection and military use. Various acoustic signal processing methods have been presented for the target detection and classification, to overcome strong ambient noise, transmission path complexity and transmission loss of underwater acoustic. Meanwhile, with the rapid development of artificial intelligence algorithms, problems in various fields are being solved by utilizing them in various fields, and among them, convolutional neural network algorithms show excellent performance in the fields of image classification. Recently, studies on the application of various artificial intelligence algorithms for vessel classification have also been actively conducted, and many studies on the application of convolutional neural networks have also been suggested for this field. However, studies on the performance of the convolutional neural network according to preprocessing in the vessel classification problem have not been confirmed. In this dissertation, it presents the performance by preprocessing in a convolutional neural network based classifier. First, it considered the vessel acoustic datasets and mentioned the preprocessing in terms of data mining and convolutional neural network that is currently widely used. Experiments began with the classification, transformation, and preprocessing of vessel acoustic data that was actually collected and released. After that, the raw data were preprocessed with feature extraction and scaling precess. In the detailed preprocessing techniques, feature extractions were applied with Mel Spectrogram and Log Mel Spectrogram techniques, and the scalers was applied with standard, min-max, max-abs, and robust techniques, respectively. In addition, exploratory data analysis was performed on the feature extraction result to understand the characteristics of the vessel acoustic data. For the experiment, a simple convolutional neural network was designed, and the classification performance of convolutional neural networks was tested for each preprocessing case by combining feature extraction and scaler techniques. For each preprocessing combination, the performance of the convolutional neural network was shown with learning curve and evaluation measures. As a result of this experiment, the performance of all scalers was improved when features were extracted by the log mel spectrogram technique. In particular, the performance of standard scaler was rapidly improved compared to when features were extracted with mel spectrogram. The combination of log mel spectrogram and robust scaler preprocessing methods showed the best classification performance. Experiment shows that preprocessing techniques affect the classification performance of convolutional neural networks, such as properly adjusting the distribution of data values while suppressing the low frequency components that act as outliers in vessel acoustic data.