Recent deep learning (DL) detectors adopted by radar or sonar (RS) are normally trained with transfer learning, where the typical workflow is to pretrain a convolutional neural network (CNN) on external large-scale classification datasets (e.g., ImageNet) as the backbone and then finetune the entire detector on detection datasets. Though transfer learning could effectively avoid overfitting, transferred models are usually redundant and might not generalize well on RS datasets. To achieve high generalization and to eliminate the dependence on transfer learning, a self-trained target detection method is established by including Automatic Deep Learning (AutoDL) to design optimal detectors. This self-trained target detection consists of three stages. First, a derived classification dataset (DCD) consisting of image blocks of targets and backgrounds is derived from detection datasets. Then, a memory-efficient Differentiable Architecture Search algorithm with flexible search space and large inputs (FL-DARTS), which is characterized by its predefined multistride convolutions, poolings, and unique super-structure, is proposed to automatically design and self-train optimal CNNs on DCDs. Finally, self-trained AutoDL detectors are implemented with the automatic backbone designed by FL-DARTS. We evaluated three self-trained AutoDL detectors on the public SAR ship detection dataset (SSDD) and the self-made sonar common target detection dataset (SCTD). The experiments show that while the number of parameters of automatic backbones designed for SSDD and SCTD are only 11.8% and 15.2% of that of ResNet50, self-trained AutoDL detectors implemented with automatic backbones significantly outperform their transfer learning detectors and achieve state-of-the-art detection precisions and high detection speeds. Data, codes are publicly available.

Underwater target detection exhibits extensive applications in marine target exploration and marine environmental monitoring. However, conventional images of underwater targets present challenges including blurred contour information, complex environmental conditions, and pronounced scattering effects. In this work, an underwater target detection method based on YOLOv10 is designed, and the detection performance is compared with the YOLOv5 model. Experimental results demonstrate that the YOLOv10 model has a mAP50 of 85.6% on the URPC 2020 dataset, improving the mAP50 by 1.2% than that of YOLOv5. This model exhibits high detection accuracy and high proceeding speed, which provides a promising support for precise and fast underwater target detection.

The utilization of polarimetric synthetic aperture radar (PolSAR) enables the preservation of more comprehensive target scattering characteristics compared to conventional SAR. Despite the significant advancements in deep learning (DL) technology for single polarimetric SAR ship detection, there remains a scarcity of DL research specifically focused on PolSAR ship detection. This article proposed a lightweight theory-driven network (LT-Net) for combining domain knowledge and DL techniques, which is suitable to detect small targets and reduce the computational complexity. The LT-Net incorporates convolutional block attention module, and employs differential downsampling during the process of high-dimensional feature extraction. The integration of domain knowledge lends each convolution step a distinct physical significance. Due to the limited availability of public datasets, this article presents a public fully polarized ship detection (FPSD) dataset for the first time, which contains 853 Pauli pseudocolor maps (in JPG format) and multilook complex data (in TIF format), with a total of 1714 ship targets from AIRSAR, UAVSAR, and RadarSAT-2. The FPSD encompasses a range of scenarios, with the majority of ship targets being small targets. Validation on FPSD shows that LT-Net exhibits significantly lower time complexity and space complexity compared to four popular target detection networks, Meanwhile, LT-Net achieves the best detection performance in the Pauli pseudocolor maps.

In multi-receiver Synthetic Aperture Sonars (SAS), the traditional non-stop-hop-stop model normally ignores the intra-pulse motion of sonar platform. However, such intra-pulse motion brings the Doppler scale effect to the echo signal, which will lead to the change of Carrier Frequency (CF) and Frequency Modulation (FM) rate of echo signal and may eventually result in image defocus. In this paper, considering the intra-pulse motion, the time-varying time delay and range history of multi-receiver SAS were derived by introducing the range time into the range history, and then the Doppler scale signal model was established. After approximating the derived time-varying range history, the two-dimensional (2D) spectrum was derived by using the shear theorem. When compared with the traditional 2D spectrum, an interesting phenomenon was found that the change of FM rate has no effect on 2D spectrum. Actually, only two additional terms, i.e., range cell migration (ARCM) and azimuth modulation (AAM), were introduced into the new 2D spectrum due to the change of CF. Besides, the condition of neglecting the ARCM was given by further analyzing the 2D spectrum, while the AAM is usually small and can be easily compensated together with the ARCM correction. Finally, a Chirp-z transform (CZT) algorithm for the Doppler scale signal model was proposed to correct the ARCM, which can simultaneously correct the ARCM and compensate the AAM (although not required) in 2D frequency domain by phase multiplication. The effectiveness of the signal model and proposed algorithm is verified by simulations and field data.

Precise orientation information facilitates target recognition of sonar images. Traditional orientation estimation methods, including the Hough transform method, have poor antinoise abilities, so they cannot achieve high estimation precision for sonar images, which usually have low signal-to-noise ratios. The convolutional neural network (CNN) regression method is sensitive to image affine transformations and requires extensive computation, so it cannot achieve strong robustness or high speed. To achieve high-precision, high-speed, and robust orientation estimation of sonar image targets, we present a novel orientation estimation method via the wavelet subimage energy ratio (WSER). The WSER varies with the rotation angles of images and has the highest value when the long axes of targets are vertical. It is translational and scale invariant and does not need supervised training. Therefore, we propose estimating orientations of sonar image targets by finding the max of the WSER to achieve high precision, high speed, and strong robustness to the translation and scale transformations of images. The results of experiments on a self-made sonar image dataset show that the mean absolute error (MAE) of the proposed method is 7.9°, while the MAEs of the CNN regression method and traditional methods are 10.1 and 22.9°, respectively. In addition, the proposed method is six times faster than the CNN regression method. The proposed orientation estimation method has also been applied to align the orientation of sonar images and CNN can achieve state-of-the-art performance for rotated target recognition using the aligned images. Data and codes are publicly available.

Representing and rendering dynamic underwater scenes present significant challenges due to the medium’s inherent properties, which result in image blurring and information ambiguity. To overcome these challenges and accomplish real-time rendering of dynamic underwater environments while maintaining efficient training and storage, we propose Underwater Dynamic Scene Reconstruction Gaussian Splatting (UDR-GS), a method based on Gaussian Splatting. By leveraging prior information from a pre-trained depth estimation model and smoothness constraints between adjacent images, our approach uses the estimated depth as a geometric prior to aid in color-based optimization, significantly reducing artifacts and improving geometric accuracy. By integrating depth guidance into the Gaussian Splatting (GS) optimization process, we achieve more precise geometric estimations. To ensure higher stability, smoothness constraints are applied between adjacent images, maintaining consistent depth for neighboring 3D points in the absence of boundary conditions. The symmetry concept is inherently applied in our method by maintaining uniform depth and color information across multiple viewpoints, which enhances the reconstruction quality and visual coherence. Using 4D Gaussian Splatting (4DGS) as a baseline, our strategy demonstrates superior performance in both RGB novel view synthesis and 3D geometric reconstruction. On average, across multiple datasets, our method shows an improvement of approximately 1.41% in PSNR and a 0.75% increase in SSIM compared with the baseline 4DGS method, significantly enhancing the visual quality and geometric fidelity of dynamic underwater scenes.