Visual Searching Method for Floating UUV Based on Deep Neural Network

With the continuous progress in unmanned underwater vehicle (UUV) recovery technologies, an increasing number of recovery tasks now rely on unmanned vessels and other autonomous systems. These devices primarily use visual detection for target localization during recovery operations. However, traditional visual detection methods often fail to accurately identify targets in complex marine environments due to interference from sea waves. This paper proposes a visual searching method based on an improved YOLOv11 model for floating UUV. A UUV target dataset is constructed to address the various interferences encountered by unmanned vehicles in different marine conditions and the varying degrees to which their bodies may be submerged. Data augmentation techniques, such as flipping, rotation, erasure, and noise addition, were applied to the constructed dataset. A lightweight image feature extraction module, the CGB module, is designed and integrated to replace the original C3k2 module in the YOLOv11 model. Experimental results show that the improved model reduces the number of parameters and inference time, achieving increases of 1.1%, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$12.1 \%, 6.1 \%$</tex>, and 9.0% in precision, recall, [email protected], and [email protected]:.95.