SEDRNet: enhanced residual networks with SE and dilated convolutions for synthetic aperture radar ship recognition

Convolutional neural networks (CNNs) are effective for classifying ship targets in synthetic aperture radar (SAR) images. However, traditional CNNs use fixed-size kernels, which can miss some important features. This often leads to inaccurate classifications. In addition, these methods tend to overlook channel information flow compared with hierarchical information flow. To address these issues, we propose a new approach called squeeze excitation dilated residual convolutional network (SEDRNet). This model reduces the number of parameters while increasing the network’s receptive field. It also learns important features from various channels. SEDRNet enhances useful features and suppresses irrelevant ones. We conducted a series of ship classification experiments using high-resolution TerraSAR-X SAR imagery. We also tested five other models: support vector machine, kernel extreme learning machine, residual neural network, densely connected convolutional network, and GoogleNet. The results show that SEDRNet outperforms these models with a classification accuracy of 99%. It also achieves strong results in other multiclass classification metrics. To validate SEDRNet’s effectiveness, we tested it on the FUSAR Ship and FGSC-23 datasets. The classification accuracies were 86.21% and 82.69%, respectively. These results are better than those from the other five methods. Finally, we compared the computational costs of SEDRNet with other models. SEDRNet takes 22.32 s to complete one epoch of training. Its inference time for a single image is 20.42 ms, which is slightly longer than DenseNet-121’s 18.54 ms.