Noise-Robust Object Detection for Forward-Looking Sonar Imagery

Because of variations in measurement conditions and marine environments, Forward-Looking Sonar (FLS) images are susceptible to uncertain noise disturbances, which can lead to degradation of FLS images and unreliable detection outcomes. To address the aforementioned issue and enhance the noise robustness of object detectors, this study proposes three models, which are DN-YOLO, SL-DN-YOLO, and FSL-DN-YOLO. Specifically, we design a Noise Injection Module (NIM) and a Denoising Module (DM), incorporating them sequentially into the backbone network of YOLOv11s for end-to-end training. The NIM is utilized only during the training phase to introduce noise into the network, which enables the DM to be trained with denoising capabilities. Furthermore, we investigate how combination methods, insertion positions, and training methods of two proposed modules affect network noise robustness. The distinctions among DN-YOLO, SL-DN-YOLO, and FSL-DN-YOLO arise from their insertion positions and training methods of NIM and DM in the YOLOv11s backbone network. DN-YOLO is employed when hardware computational capacity is limited and faced with low noise levels, whereas SL-DN-YOLO is used when there are significant noise levels. FSL-DN-YOLO is applicable when adequate computational capacity is available and dealing with noise disturbances. Experimental results confirm that all three proposed models exhibit strong resistance to noise disturbances.