基于时域–频域协同引导的海洋机器人视觉感知增强方法

Time-frequency Collaborative Guidance-based Visual Perception Enhancement for Marine Robots

摘要:

在海洋机器人执行水下探测与作业任务过程中,受水体对光线吸收与散射效应的影响,采集到的水下图像普遍存在亮度不足、细节模糊及色彩失真等问题,严重制约了后续视觉感知与智能决策任务的性能。针对现有方法多侧重于时域信息建模而忽略频域特征在结构保持与细节恢复中重要作用的不足,提出了一种时域–频域协同引导的海洋机器人视觉感知增强模型。从整体结构上看,所提出的模型主要包括时域–频域特征提取与多维度特征细化两个阶段。在时域–频域特征提取阶段,分别构建时域分支与频域分支并引入跨域协同模块实现多域特征的交互融合,从而增强对水下图像结构与纹理信息的综合表征能力。在多维度特征细化阶段,进一步设计多种特征增强模块以提升重建质量,首先提出偏振残差模块,通过多方向特征提取与自适应融合,在突出有效结构信息的同时抑制噪声干扰,实现细节与对比度的协同增强;其次,构建双注意力耦合模块,联合通道注意力与空间注意力机制,在保持全局结构一致性的基础上强化局部关键细节;最后,设计双重感受野的多尺度残差模块,结合多尺度卷积、注意力调节及可学习缩放参数,实现对关键特征的自适应强调,在细节保真与整体视觉效果之间取得平衡。在多个公开水下数据集上的实验结果表明,本文方法在主观视觉效果上均优于现有主流视觉增强方法,并在复杂水下环境中表现出更好的鲁棒性与泛化能力,为海洋机器人视觉感知与下游任务提供有效支撑。

Abstract:

Marine robots often capture severely degraded images during underwater exploration and operation tasks due to light absorption and scattering in water, which leads to low brightness, blurred details, and significant color distortion. These degradations substantially hinder subsequent visual perception and intelligent decision tasks. To solve the limitations of existing approaches that mainly rely on time-domain modeling and neglect the role of frequency-domain features in structure preservation and detail recovery, the temporal-frequency collaborative guided network was proposed for marine robots. The proposed network consisted of two stages: time-frequency feature extraction and multi-dimensional feature refinement. In the time-frequency feature extraction stage, a time domain branch and a frequency domain branch were employed to capture complementary representations, and a cross domain collaboration module was introduced to explicitly fuse multi domain features, thereby strengthening the representation of underwater image structures and textures. In the multi-dimensional feature refinement stage, several enhancement modules were incorporated to further improve reconstruction quality. Specifically, polarization residual module extracted multi-directional features and performed adaptive fusion were proposed to emphasize effective structural information while suppressing noise, which enabled simultaneous enhancement of image details and contrast. In addition, dual attention coupling module combined channel attention and spatial attention was introduced to reinforce local salient details while preserving global structural consistency. Finally, multi-scale residual module with dual receptive fields integrated multi-scale convolutions, attention modulation, and learnable scaling parameters was proposed to adaptively highlight critical features and balance fine-detail preservation with overall visual quality. Extensive experiments on multiple public underwater image datasets showed that the proposed method consistently outperformed state-of-the-art visual enhancement approaches in subjective visual quality, and demonstrated strong robustness and generalization capability in complex underwater environments, effectively supporting marine robot visual perception and downstream tasks.