Aiming at the problem of small target recognition caused by low resolution, low signal-to-noise ratio (SNR) and insufficient multi-scale feature fusion in underwater acoustic imaging, this paper proposes a multi-scale feature reconstruction and accurate recognition algorithm (MSFF-Net). The algorithm adopts the encoder-decoder architecture, and designs a dynamic weight fusion (DWF) strategy to adaptively reconstruct the feature pyramid, so as to realize the efficient fusion of high-level semantics and low-level details. Differential Attention (DA) mechanism is introduced to enhance the target edge and suppress the background noise through cross-scale feature differences. Focal Loss is combined to alleviate the problem of sample imbalance, and channel pruning and depth-separable convolution are used to realize the model lightweight. Experiments on SMD data sets show that MSFF-Net's small target detection accuracy (APₛmall) is 46. 7%, and email protected is 76. 8%, which is definitely 6. 6% higher than mainstream methods such as YOLOv5s, while the parameters are reduced to 5. 8M and FLOPs is only 14. 3G. Ablation experiments verify the complementarity of DWF and DA mechanism and the remarkable enhancement of DA to small target features, and cross-domain tests prove its excellent generalization ability. The algorithm can effectively improve the detection accuracy and robustness of small targets in complex underwater environment while maintaining light weight.
Wu et al. (Sun,) studied this question.