Deep learning is a core technology for sonar image classification. However, owing to the cost of sampling, a lack of data for sonar image classification impedes the training and deployment of classifiers. Classic deep learning models such as AlexNet, VGG, GoogleNet, and ResNet suffer from low recognition rates and overfitting. This paper proposes a novel network (ResNet-ACW) based on a residual network and a combined few-shot strategy, which is derived from generative adversarial networks (GAN) and transfer learning (TL). We establish a sonar image dataset of six-category targets, which are formed by sidescan sonar, forward-looking sonar, and three-dimensional imaging sonar. The training process of ResNet-ACW on the sonar image dataset is more stable and the classification accuracy is also improved through an asymmetric convolution and a designed network structure. We design a novel GAN (LN-PGAN) that can generate images more efficiently to enhance our dataset and fine-tune ResNet-ACW pretrained on mini-ImageNet. Our method achieves 95.93% accuracy and a 14.19% increase in the six-category target sonar image classification tasks.

Detecting small undersea and buried targets, such as unexploded ordnances and sunken unmanned vehicles, is vital for protecting marine ecosystems and ensuring waterway security. Challenges arise from the spatial variability of acoustic scattering from these targets, coupled with significant reverberation interference from sediment layers, which reduces the signal-to-reverberation ratio (SRR) and hinders search efficiency. To address these issues, we propose an advanced high-resolution two-dimensional (2D) spatiotemporal multi-input multi-output (MIMO) detection method designed explicitly for buried targets. This approach employs a compact low-frequency MIMO array and an orthogonal frequency-modulated pulse train signal crafted under sonar-environment constraints. To enhance the target irradiation range of our method, a 2D point spread function is developed that aligns with the acoustic scattering characteristics of cylindrical targets while effectively suppressing reverberation interference. Experimental results demonstrate that the proposed method can double the target irradiation range, enhance the SRR by over 12 dB, and increase detection probability by at least 37% compared with conventional detection techniques.

Wideband signals exhibit significant advantages in reverberation suppression, range resolution, and target parameter estimation. Consequently, echoes acquired through wideband active sonar provide rich target information, facilitating tasks such as target detection and classification. However, as range resolution increases, target echoes spread across multiple range bins, resulting in energy accumulation loss and a reduced signal-to-reverberation ratio (SRR) within individual range cells, which ultimately degrades detection probability. To address this challenge, this paper proposes a detection method based on inter-frame mutual information (IFMI) in the range-Doppler (R-D) spatial domain. Received echo signals are segmented into frames and mapped into the R-D domain, forming multi-frame R-D subspace matrices. IFMI is introduced to quantify data association states between consecutive frames. This is achieved by computing the mutual information between corresponding rows of adjacent R-D subspace matrices. The minimum IFMI is designed as the final test statistic and compared against a decision threshold to achieve target detection. Furthermore, the proposed method preserves target information within the R-D domain. The extremum positions of the IFMI can be utilized to estimate target parameters, thus avoiding extensive two-dimensional searches in the R-D domain. The efficacy and superiority of the proposed method were validated through simulations and experimental analyses. The simulation results demonstrate that the proposed method significantly improves the detection performance for extended targets under low SRR conditions, achieving a detection probability of 85% even at -12dB SRR. Experimental results further verify that, compared to four existing methods, the proposed method increases the detection probability by at least 17% and reduces the false alarm probability by 10%.

The design of transmitting waveforms is an effective way to improve the detection performance of sonar systems. For the problem of high-range sidelobe when designing reverberation-resistant waveforms, this paper proposes a high-resolution wideband composite waveform design with reverberation suppression performance and a waveform parameter improvement method. Firstly, we propose a novel wideband waveform, which utilizes linear frequency modulation (LFM) as the fundamental pulse, referred to as multi-parameter coded modulation LFM pulse (MPCM-LFM). Additionally, we deduce the wideband ambiguity function for waveform design. Then, we deduce the constraint relations of the waveform parameters for different sub-band overlaps, and according to the mathematical expressions of the obtained range ambiguity function, we analyze in detail the effects of the waveform parameters on the range ambiguity function under different constraints. Secondly, on the basis of the analysis, we also propose a hopping carrier frequency constraint rule to optimize the spectral performance, and the range sidelobe is restrained effectively in significant measure by this parameter improvement method. Finally, we analyze the computer simulation results. It is obvious that our proposed waveform parameter improvement method leads to good results. The proposed improved MPCM-LFM signal shows a “near-thumbtack” ambiguity function, whose sidelobe suppression performance is superior to other classical waveforms in the desired region, and it can realize high-precision parameter estimation. In addition, the proposed improved MPCM-LFM signal possesses good performance in detecting stationary and low Doppler targets in the background of reverberation.