Abstract: The issue of underwater litter and its associated hazards has attracted global attention, while the advancement of underwater robotics and object detection technology offering potential for the automated management of underwater debris. However, due to the high collection costs and difficulty of underwater data, applying deep learning methods to these tasks often results in a few-shot training environment, leading to a high risk of model overfitting. At the same time, the specificity of the underwater environment makes generic object detection algorithms are not well suited and require targeted improvements. Given the aforementioned two challenges, an underwater debris object detection algorithm is proposed suitable for few shot environments. Firstly, underwater data have singular foreground and a large amount of redundant noise in the background, to effectively preserve valuable information, features are extracted from the support set in focus part, enabling the model to focus more on the object itself while retaining appropriate contextual information. Secondly, to augment the model’s ability to extract information from the support set and maintain its generalization, a noise generator is introduced to send random perturbations to the focus region. Finally, considering that the support set and query set come from the same sampling domain, a joint meta-loss is proposed to make the model aware of this commonality, thereby enriching the information provided by the support set. Additionally, a diverse and contextually relevant underwater debris dataset is created, aligning more closely with real-world detection scenarios. The proposed approach achieves a precision of 16.9% under one-shot condition on this dataset, marking a 4.5?percentage points improvement over baseline models. Moreover, an increase of over 10 percentage points on the general dataset PASCAL VOC validates its generalizability.

Key words: underwater debris, few-shot learning, object detection, two-stage network, Meta R-CNN, meta learning

摘要： 水下垃圾及其危害已引起全球性的关注，而水下机器人与目标检测技术的发展为水下垃圾的自动化处理带来了可能性。由于水下数据采集成本高、难度大，难以获取充足样本，为深度学习方法进行水下垃圾检测带来困难。因此提出一种适配水下环境的小样本目标检测算法。水下数据前景单一且背景存在大量冗余噪声，为更好保留有效信息，提出聚焦式支持集处理方式，使模型能在保留适当上下文信息的同时更多关注目标本身。为进一步增强模型对支持集的信息提取能力并维持模型的泛化性，引入噪声生成器对聚焦区域添加随机扰动。考虑到支持集与查询集来自同一样本分布，提出联合元损失促使模型学习到该共性，从而丰富支持集提供的信息。此外，制作了类别多样并更贴合实际检测场景的水下垃圾数据集，所提出方法在该数据集上以单样本条件实现16.9%的精度，相较基准模型提高4.5个百分点。同时，在通用数据集PASCAL VOC上达到10个百分点以上涨幅，验证了其泛化性。

关键词: 水下垃圾, 小样本学习, 目标检测, 二阶段网络, Meta R-CNN, 元学习