Echo Highlight Image Recognition Using Zero-Shot Learning in Active Sonar Systems

Active sonar systems are pivotal in undersea operations, particularly for identifying and classifying underwater targets amidst varying disturbance levels. Traditional methodologies often struggle with the accurate discrimination of real moving targets due to the complexity introduced by dynamic underwater environments. This research introduces an enhanced target highlight model and a novel image recognition technique using zero-shot learning to effectively isolate and recognize echo highlight images from underwater targets. We developed a specialized classification convolutional neural network, termed HasNet-5, engineered to extract both global features and specific local highlight characteristics of these images. HasNet-5 operates on a unique framework of zero-shot learning, enabling it to classify images without prior direct exposure to specific target data during training. Utilizing five distinct target highlight models, the system was trained, validated, and tested through both simulated and experimental data. The empirical evaluation demonstrated that our approach achieved a recognition accuracy of 92 % for real targets and 94 % for identifying two-dimensional disturbance targets. The results substantiate the efficacy of integrating zero-shot learning with active sonar technology, presenting a significant advancement in underwater target recognition capabilities.