Synthetic Aperture Radar (SAR) is limited by scattering characteristics, wavelength interference, and other factors when detecting ship targets, making it difficult to recognize targets and obtain boundary information, which affects detection accuracy. In order to improve SAR ship detection accuracy while reducing false detection rates, an SAR ship detection method is proposed based on a stepped residual structure and coordinate information recombination. First, in the backbone network, as the main feature extraction module for improving ship target recognition ability and effectively reducing parameter quantity, a sequential residual convolution block is constructed. Second, in the feature fusion part, a spatial channel attention mechanism and a cascaded layer are used to construct a block attention network, further focusing on the geometric information of the model in the detection head part. The coordinate information reconstruction convolution is proposed in the final stage of feature transfer. The characteristics of coordinate information recombination convolution rotation invariance can be well utilized to improve the inspection effect of ships in different directions at sea while enhancing the model's ability to bear interference, optimizing the quality of information fusion. Finally, the normalized Gaussian Wasserstein Distance (NWD) is introduced into the loss regression function at the detection head to enhance the detection ability of small targets.

First, a sequential residual partial (SRP) convolution block is constructed in the backbone network. It uses depthwise separable convolution, partial convolution, and regular convolution to perform multi-scale feature extraction through a multi-residual sequential fusion and ladder-like dense connection method, effectively reducing the number of parameters. Second, a spatial channel attention mechanism and cascaded layers are used to build a block attention network (SCAA-Net) in the feature fusion part. By replacing the ReLU activation function with Leaky ReLU and integrating features from different network depths, we enhance the ability to extract geometric features and details. Then, coordinate information reconstruction convolution (CIRConv) is proposed at the end of the feature transfer. It uses fractional Fourier transform to capture time-frequency characteristics of signals and provide more discriminative basis for the detection algorithm. Finally, the normalized NWD is introduced into the loss regression function of the detection head to enhance the detection ability of small targets.

To demonstrate the effectiveness of the proposed network, experiments are conducted on the HRSID dataset and SSDD dataset. The results show that for more complex HRSID datasets, in comparison with the benchmark model, the accuracy of the stepped residual and coordinate information reconstruction network increased by 4.8%, the recall rate increased by 3.7%, the average accuracy increased by 4.1% when Intersection over Union (IoU) was 0.5, and the IoU increased by 4.1%. The average accuracy increased by 2.7% in the SSDD dataset when IoU was 0.5∶0.95. Compared with the baseline model, the accuracy of the hierarchical residual and coordinate information recombination network increased by 4.9%, the recall increased by 1.3%, the average accuracy increased by 2.8% when IoU was 0.5, and the average accuracy increased by 2.8% when IoU was 0.5∶0.95. The network model parameters reduced by about 20%. The proposed network has significant advantages in improving SAR ship detection accuracy, improving false positives and omissions, and effectively solving the problem of SAR ship detection,which provides an effective and high-precision method for detecting ships.

The proposed SAR ship detection method based on ladder residual and coordinate information recombination effectively improves the detection accuracy, reduces the false detection rate, and has better performance than mainstream algorithms. The innovative modules in the network play important roles in feature extraction, fusion, and loss function optimization, providing an effective and high-precision solution for SAR ship detection.