Task-driven joint denoising and super-resolution of side-scan sonar images based on knowledge distillation

Denoising and super-resolution of side-scan sonar (SSS) images are crucial for high-level computer vision tasks such as underwater object detection. However, treating denoising and super-resolution independently can cause the two processes to interfere with each other. Moreover, improved visual quality does not necessarily yield better downstream task performance. Achieving task-driven joint denoising and super-resolution of SSS images remains a challenge. Focusing on detection tasks, this paper proposes a task-driven joint denoising and super-resolution method for SSS images based on knowledge distillation. First, a multi-factor degradation module synthesizes training data that approximate real-world degradations. Second, a task performance tracking module compares the task responses produced by the student output and the ground-truth image, guiding the student toward improved downstream performance. Third, an image detail tracking module combines the visual characteristics of SSS images with task requirements and enforces consistent regional discrimination between the student and teacher outputs. Finally, a negative sample generation and filtering module uses selected negative samples to impose a lower bound constraint on the student output, thereby simplifying optimization of the solution space. Experimental results show that the proposed method outperforms comparison methods in quantitative metrics and provides effective assistance for high-level vision tasks.