In recent years, the rapid development of Deep Learning (DL) has provided a new method for ship detection in Synthetic Aperture Radar (SAR) images. However, there are still four challenges in this task. (1) The ship targets in SAR images are very sparse. A large number of unnecessary anchor boxes may be generated on the feature map when using traditional anchor-based detection models, which could greatly increase the amount of computation and make it difficult to achieve real-time rapid detection. (2) The size of the ship targets in SAR images is relatively small. Most of the detection methods have poor performance on small ships in large scenes. (3) The terrestrial background in SAR images is very complicated. Ship targets are susceptible to interference from complex backgrounds, and there are serious false detections and missed detections. (4) The ship targets in SAR images are characterized by a large aspect ratio, arbitrary direction and dense arrangement. Traditional horizontal box detection can cause non-target areas to interfere with the extraction of ship features, and it is difficult to accurately express the length, width and axial information of ship targets. To solve these problems, we propose an effective lightweight anchor-free detector called R-Centernet+ in the paper. Its features are as follows: the Convolutional Block Attention Module (CBAM) is introduced to the backbone network to improve the focusing ability on small ships; the Foreground Enhance Module (FEM) is used to introduce foreground information to reduce the interference of the complex background; the detection head that can output the ship angle map is designed to realize the rotation detection of ship targets. To verify the validity of the proposed model in this paper, experiments are performed on two public SAR image datasets, i.e., SAR Ship Detection Dataset (SSDD) and AIR-SARShip. The results show that the proposed R-Centernet+ detector can detect both inshore and offshore ships with higher accuracy than traditional models with an average precision of 95.11% on SSDD and 84.89% on AIR-SARShip, and the detection speed is quite fast with 33 frames per second.

Deep learning adjusts parameters to optimize the model through model training. Training algorithm is the key to model optimization and implementation. Therefore, the improvement of model training algorithm is of great significance to deep learning. Based on the original gradient algorithm, the paper proposes a new gradient descent algorithm AdaGrad Restricted by Windows (AdaRW) for Deep learning model training optimization. Aiming at the defects of AdaGrad algorithm, the new algorithm uses the subset of window to limit historical accumulation, so as to slow down the attenuation of learning rate, and improve the speed of model training. The paper constructs OceanTDA9, a Deep learning model of marine target detection for Synthetic Aperture Radar (SAR) data, and adopts the proposed AdaRW algorithm to train the model based on SAR data with a resolution of 10 m in the Bohai Sea. Experiment shows that the accuracy and loss of the algorithm are better than those of AdaGrad and Stochastic Gradient Descent (SGD) algorithms, and the standard deviation is better than that of Adam algorithm.

This study constructs four deep-learning OceanTDLx series models and uses a WinR-AdaGrad gradient descent algorithm to train and optimize the constructed models. Through an analysis of the loss, accuracy, and time consumption of the four models (i.e., OceanTDL2, OceanTDL3, OceanTDL5 and OceanTDL8), we reveal that the models’ performance does not improve when the number of layers is increased and that OceanTDL5 provides the optimal performance. OceanTDL5 is compared with OceanTDA9 (a model that we previously constructed), and the curves for training loss_batch and training accuracy_batch indicate that OceanTDL5 is more suitable than OceanTDA9 for detecting distributed targets, particularly semi-melted sea ice, which is intertwined and easily confused with seawater. We process the SAR (Synthetic Aperture Radar) data of the research area and obtain a data set with a 10-m resolution, which is then used to verify the effectiveness of the constructed models for sea ice detection. The results reveal that OceanTDL5 has a detection capacity of approximately 55.6 km2 /s and a detection accuracy rate of 97.5%. Compared with traditional ocean target detection methods, OceanTDL5 has greater detection speed and accuracy.