中国科学院自动化研究所机构知识库

       海洋蕴含丰富的生物与矿产等资源，可靠、高效的海洋探索对人类发展具有重要意义。本文围绕机器人水下自主探索中的视觉感知问题，从小样本目标检测、开集目标检测以及增量目标检测三个方面逐层递进展开深入研究，旨在为水下机器人在海洋环境下的智能自主探索提供理论指导和技术支撑。取得的主要研究成果如下：

       一、针对小样本目标检测中特征耦合以及分类头容易过拟合的问题，提出了一种解耦度量的单阶段小样本目标检测方法。首先，构建了一种解耦表达转换模块（Decoupled Representation Transformation，DRT），提出了特征自适应变换方法，实现了分类定位特征解耦，提升了检测器在新类上的泛化性。其次，提出了一种图像级度量学习模块（Image-Level Distance Metric Learning，IDML），实现了对整张特征图的直接处理，无需感兴趣区域（Region of Interest，RoI）池化操作，适用于单阶段目标检测。得益于度量学习较高的泛化性，所提IDML进一步提升了检测器对新类的检测精度。最后，结合所提的DRT与IDML，构建了解耦度量网络（Decoupled Metric Network，DMNet），实现了较高性能的单阶段小样本目标检测。大量实验验证了所提方法的有效性、实时性与鲁棒性。水下目标检测实验结果表明，DMNet在真实水下场景能够很好地完成水下小样本目标检测任务。

       二、针对小样本检测中多分类范式与小样本任务间的冲突以及困难负样本问题，提出了一种能够抑制困难负样本的高精度两阶段小样本目标检测方法。首先，从小样本目标检测出发，得出多重二分类范式比现有多分类范式更适合小样本检测任务的结论。其次，基于多重二分类范式，设计了一种二元相似度头，在新类上具有更高的泛化性。再次，针对困难负样本设计了一种特征增强模块，加大了正样本与困难负样本特征间的差异，降低了假阳性率。最后，基于上述所提模块，提出了一种二元相似度检测器（Binary Similarity Detector，BSDet），实现了较高的小样本目标检测精度。大量实验验证了所提方法的有效性和优越性，表明多重二分类范式更适合小样本目标检测任务。野外湖泊场景实验结果表明，BSDet在真实水下场景能够很好地完成水下小样本目标检测任务。

       三、针对开集目标检测问题，提出了一种基于分类定位融合的开集目标检测方法。首先，结合分类与定位的特点，提出了一种融合区域建议网络，实现了对未知类物体泛化性更好的前景置信度预测。其次，提出了一种未知类伪标签选取方法，实现了训练过程中对未知类物体的自动标注，避免了潜在的前景物体被分配为背景标签，损害对未知类物体的检测性能。再次，提出了面向已知类和未知类物体的差异增强模块，提高了检测器区分已知类与未知类物体的能力，缓解了两者混淆的问题。最后，大量实验和可视化验证了所提方法的有效性，在公开数据集上与野外湖泊场景均较好地完成了对未知类物体的检测。

       四、针对机器人水下自主探索任务，提出了一套完整的视觉感知框架。首先，梳理并规范化了机器人水下自主探索任务，将其拆分为闭集目标检测、开集目标检测以及增量目标检测，建立了三者的相互关系。其次，提出了一种余弦相似度聚类，实现了无监督情况下自适应地确定聚类簇数，以及不同未知类物体之间的细分。再次，提出了一种回传梯度缩放策略以及知识蒸馏方法，使检测器在增量学习时能较好保留旧类知识。最后，提出了一套完整的视觉感知框架，实现了机器人的水下自主探索，完成了已知类物体检测、未知类物体检测、未知类物体细分以及对新类物体的增量学习等任务。大量实验结果表明了所提方法的有效性，机器人在野外水下场景能够有效完成自主视觉感知任务。

       The ocean is rich in biological and mineral resources. Reliable and efficient ocean exploration is of great significance to human development. This dissertation focuses on the visual perception in underwater autonomous exploration of robots. To generate fresh insights into underwater robotics autonomous exploration in marine environments, it carried out research work from three key aspects, including few-shot object detection, open-set detection and incremental object detection. The main contributions are summarized as follows.

       Firstly, a decoupled metric single-stage few shot object detection method is proposed to solve the problems of feature coupling and overfitting of the classification head in few-shot obejct detection. As for the problem of feature coupling, a decoupled representation transformation (DRT) is proposed. The DRT executes a feature adaptive transformation to obtain decoupled classification and localization features and improves the generalization to novel classes. Thereafter, to alleviate the overfitting of the classification head, an image-level distance metric learning (IDML) is proposed. The IDML directly processes the whole feature map without RoI pooling, so it is suitable for single-stage object detection. Benefitted from the high generalization of metric learning, the proposed IDML further improves the detection precision of the detector for novel classes. Further, combining the proposed DRT and IDML, a decoupled metric network (DMNet) is built to realize high-performance single-stage few-shot object detection. Extensive experiments verify the effectiveness, real-time and robustness of the proposed method. Especially, the experimental results of underwater object detection show that the DMNet can perform the underwater few-shot object detection well in real underwater scenarios.

       Secondly, a high-precision two-stage few-shot object detection method is proposed to alleviate the conflict between multi-classification paradigm and few-shot tasks and suppress hard negative samples. From the perspective of few-shot object detection, we draw a conclusion that the multiple binary classification paradigm is more suitable for the few-shot detection, compared to the existing multi-classification paradigm. Then, based on the multiple binary classification paradigm, a binary similarity classification head is exploited, achieving higher generalization over novel classes. Besides, a feature enhancement module is designed for hard negative samples, which increases the difference between the features of positive and hard negative sample and reduces the false positive rate. Based on the above proposed module, a binary similarity detector (BSDet) is proposed, achieving high-precision few-shot object detection. A large number of experiments verify the effectiveness and superiority of the proposed method, indicating that the multiple binary classification paradigm is more suitable for few-shot object detection. Furthermore, the experimental results in a field lake scenario show that BSDet can also perform well in real underwater scenarios.

       Thirdly, a classification and localization fusion-based open-set object detection method is proposed for the open-set object detection. Combining the characteristics of classification and localization, a fusion region proposal network is proposed, achieving better objectness prediction for unknown objects. Further, an unknown class pseudo-label selection method is proposed to automatically label unknown objects during the training. The UPLS can avoid potential foreground objects being assigned as background, which could harm the detection performance of unknown objects. Moreover, for known and unknown object detection problems, a difference enhancement module (DEM) is proposed to improve the ability of detector to distinguish known and unknown objects and alleviate their confusion. Extensive experiments and visualizations demonstrates the effectiveness of the proposed method, achieving good detection performance on both public datasets and field lake for unknown objects.

       Fourthly, a complete visual perception framework is proposed for robotic underwater autonomous exploration. The robotic underwater autonomous exploration is sorted out and normalized. As a result, it is divided into closed-set object detection, open-set object detection, and incremental object detection, and the interconnection of the three is established. Then, a cosine similarity clustering is proposed to adaptively determine the number of clusters and to subclassify different unknown objects in an unsupervised manner. Besides, to enable the detector preserve old knowledge during incremental learning, a back-propagation gradient scaling strategy and a knowledge distillation method are proposed. Further, a complete visual perception framework is proposed and completes robotic autonomous underwater exploration. Specifically, it successfully achieves many objects, including known object detection, unknown object detection, unknown object subdivision, and incremental learning of novel class objects. Numerous experimental results demonstrate the effectiveness of the proposed method, and the robot can effectively complete the autonomous exploration in the field underwater scene.