An improved super-resolution GAN for feature enhancement in surface ship target detection

Maritime remote sensing images acquired by satellites and unmanned aerial vehicles (UAVs) often exhibit low resolution, loss of fine details, strong background interference, and arbitrary target orientations. Detecting small-scale, densely distributed ships, particularly in coastal and port regions with high ship density, remains a critical and challenging problem. To address these challenges, we propose a novel super-resolution reconstruction network, Wasserstein–Attention–Super–Resolution (WASR), based on the Super-Resolution Generative Adversarial Network (SRGAN). This approach is designed to to strengthen the feature representation of small-scale and densely distributed ship targets in remote sensing images. In the generator design, we introduce a Residual–Squeeze-and-Excitation (Residual-SE) network to improve feature extraction capabilities. In the discriminator, both the convolutional blocks and the first fully connected layer utilize the mish activation function. Additionally, a gradient penalty term is added during training to redefine the loss function, yielding more stable optimization.Experimental results demonstrate that with the same YOLOv5 detection model, compared with the original image, SRResNet, SRGAN, and Real-ESRGAN methods, the AP50 of the images enhanced by WASR is increased by 3.8%, 4.1%, 2.3%, 2.0%, respectively, while the AP50 : 95 is improved by 8.1%, 8.9%, 5.1%, 3.6%, respectively. Meanwhile, the average detection time per image is reduced by 9.40%, 1.67%, 7.81%, and 14.34%, respectively.