A small target detection method for undersea sonar imagery based on super-resolution reconstruction

Small-target detection in underwater sensing systems is particularly challenging. For the side-scan sonar imagery, the problem is aggravated by low spatial resolution and strong acoustic noise. Speckle noise introduces spurious highlights around targets, and ambient noise can mask weak acoustic backscatter. These factors lead to missed detections and false alarms. The problem becomes more severe after backbone downsampling, and small targets may occupy only a few pixels. Although recent deep-learning approaches have explored super-resolution techniques to enhance image quality, existing methods often fail to achieve a suitable trade-off between detection accuracy and computational efficiency. This paper proposes a target detection model named Impro-YOLO-S for small-target detection in low-resolution imageries. Impro-YOLO-S is composed of a Perceptual and Frequency Loss Enhanced SRGAN and a Re-parameterized YOLO detection network. The SRGAN module includes a global-local receptive field loss enhanced by a Swin-transformer branch and a ResNet branch. Additionally, this module integrates a fast Fourier transform frequency loss to improve frequency domain representation. Specifically, we compute a frequency-domain loss via fast Fourier transform (FFT) and apply inverse FFT to project the spectral discrepancy back to the image space. Impro-YOLO-S is trained end-to-end, and the detection loss is backpropagated to the SR generator to drive task-relevant reconstruction and alleviate SR–detector feature mismatch. The Re-parameterized YOLO detection network employs a structural optimization strategy that mathematically fuses convolution and batch normalization parameters into YOLO detection network. The experimental results show that Impro-YOLO-S achieves an mAP of 74.45%, which represents improvements of 14.37%, 9.93%, 57.87%, and 68.36% compared to YOLOv8, YOLOv5, RetinaNet, and Faster R-CNN, respectively. We also compare with wavelet energy-based methods KWE and SWE. Compared with wavelet energy-based methods, Impro-YOLO-S generates more accurate and stable bounding boxes with correct semantic labels across challenging scenarios. Moreover, the GFLOPS of Impro-YOLO-S reaches 39.9, thus meeting the requirements for real-time search.