Sonar is the primary tool used for detecting small targets at long distances underwater. Due to the influence of the underwater environment and imaging mechanisms, sonar images face challenges such as a small number of target pixels, insufficient data samples, and uneven category distribution. Existing target detection methods are unable to effectively extract features from sonar images, leading to high false positive rates and affecting the accuracy of target detection models. To counter these challenges, this paper presents a novel sonar small-target detection framework named DCE-Net that refines the YOLOv8 architecture. The Detail Enhancement Attention Block (DEAB) utilizes multi-scale residual structures and channel attention mechanism (AM) to achieve image defogging and small-target structure completion. The lightweight spatial variation convolution module (CoordGate) reduces false detections in complex backgrounds through dynamic position-aware convolution kernels. The improved efficient multi-scale AM (MH-EMA) performs scale-adaptive feature reweighting and combines cross-dimensional interaction strategies to enhance pixel-level feature representation. Experiments on a self-built sonar small-target detection dataset show that DCE-Net achieves an mAP@0.5 of 87.3% and an mAP@0.5:0.95 of 41.6%, representing improvements of 5.5% and 7.7%, respectively, over the baseline YOLOv8. This demonstrates that DCE-Net provides an efficient solution for underwater detection tasks.
The quality of sonar images is crucial for underwater perception and interpretation. However, due to the low resolution, weak separability, and strong interference characteristics of sonar data, improving the quality of sonar images has always been a highly challenging task. Existing super-resolution technologies, such as Real-enhanced super-resolution generative adversarial network (ESRGAN), although they show certain effects in image reconstruction, still fail to meet the strict requirements for sonar image processing. They often lead to artifacts and oversmoothing in the reconstructed images, which degrade image fidelity and limit further analysis. Aiming at the limitations of Real-ESRGAN in reconstructing underwater sonar images, this letter proposes a sonar image reconstruction method based on enhanced Real-ESRGAN. During the sonar image reconstruction process, this method employs a noise reduction treatment that combines deep learning-based nonlocal means filtering with adaptive wavelet transformation. This effectively reduces artifacts caused by aliasing effects. Moreover, the CoordGate module is introduced into the backbone network residual in residual dense block network of Real-ESRGAN, significantly enhancing the model's feature modulation capability. Experimental results indicate that this method performs excellently in sonar image reconstruction, particularly in terms of peak signal-to-noise ratio and structural similarity index measure, producing visually sharper and structurally more consistent results compared to the baseline method.
Synthetic aperture sonar is the main means of underwater long-range and high-resolution imaging, and small-sample classification based on synthetic aperture sonar images holds important application value. However, due to the constraints of sonar imaging mechanism, acquisition cost and other factors, synthetic aperture sonar images face the problems of insufficient target pixels and scarce sample size, which limit the performance of existing target classification methods. In this paper, super-resolution reconstruction and contrastive learning mechanism are used to achieve highly accurate classification of synthetic aperture sonar images with small samples. Considering the spatial distribution characteristics of fuzzy images, the super-resolution reconstruction algorithm is introduced to realize the pixel expansion and structural complementation of blurry images. To address the problem of scarce samples, a small-sample classification mechanism based on contrastive learning is innovatively designed, achieving high-accuracy classification of small-sample sonar images post super-resolution reconstruction. The results of the classification experiments on the SCTD dataset show that the proposed method can improve the accuracy by 4.17% compared with the classical classification method based on convolutional neural network. The proposed method can not only serve to improve the performance of existing small-sample sonar images classification methods, but also illustrate the applicability of contrastive learning in the field of sonar image classification.