结合二值化神经网络与知识蒸馏的轻量型水声目标识别算法

Underwater acoustic target recognition (UATR) algorithms based on deep learning often face the challenges of scarce computing resources and the complex and variable underwater acoustic environment when deployed on underwater platforms. Therefore, this paper proposes a lightweight UATR algorithm depthwise separable binarized neural network with knowledge distillation (DSBNN_KD), so as to achieve model compression and optimized acceleration through means of depth-separable convolution and weight parameter binarization. Meanwhile, the KD is utilized to transfer knowledge from high-performance, high-complexity teacher models to lightweight student models, thereby mitigating the performance loss caused by extreme quantization and ensuring the model’s generalization performance. The performance of the DSBNN_KD is comprehensively evaluated on two observed underwater acoustic datasets. The experimental results indicate that, compared to current mainstream full-precision lightweight models, the proposed DSBNN_KD shows significant advantages in terms of model parameter volume, model deployment size, and computational load. With the assistance of the KD, the quantized model can still maintain performance close to that of full-precision models.