소량 불균형 데이터셋의 거리 및 방위 특징을 활용한 딥러닝 모델 기반 단상태 능동 소나 표적 식별 시스템 연구

This paper proposes a system for a automatic active sonar target classification system that utilizes target detection images in the bearing-range domain as input feature. These images are obtained by applying conventional active sonar signal processing techniquis such as beam-forming and matched filtering to the received signals. Deep Convolutional Neural Networks (DCNN) and Convolutional Recurrent Neural Network (CRNN) were employed as classifiers, and their target classification performance was compared with respect to the proposed input features. Two simulation datasets were generated using a simulator under a monostatic active sonar scenario, each assuming different ocean environments, and were used for model training. The performance of the proposed system was evaluated on validation datasets using accuracy as well as target and non-target F1-score as metrics. The experimental results demonstrate that the proposed features, which incorporate both bearing and long-range detection information, can be effectively exploited by neural network architectures regardless of the transmitted pulse type. Also, the meaningful target classification performance can be achieved even with a limited amount of imbalanced training data without the application of additional data augmentation.