毕业生知识库

遥感技术是一种非接触的、远距离探测技术，可以通过搭载在特定平台上的传感器或遥感器，获取不同波段、不同分辨率的信息。其中，光学遥感技术占据重要的地位。随着遥感平台技术的发展，光学遥感影像数据爆炸性增长。如何高效、准确地解译这些数据，成为了遥感领域亟待解决的问题。在这一背景下，目标检测技术作为遥感图像解译的基石，发挥着至关重要的作用。它不仅能够准确识别图像中特定目标的位置，还能对这些目标进行精细分类，进而为城市规划、环境监测、资源管理等应用提供坚实的数据支撑。因此，开展光学遥感图像目标检测的研究，具有重要的理论和实际应用价值。近年来，光学遥感图像目标检测的研究经历了从传统方法到深度学习方法的重要转变，并在实时性和准确性等方面实现了突破。然而，随着应用场景的日益复杂化和专业化，标注受限问题逐渐凸显，成为了制约目标检测性能提升的关键因素。首先，标注信息不充足导致模型性能不能满足应用场景对高精度的需求；其次，标注信息不确切导致模型学习到错误的特征表示。因此，本文立足于标注信息受限的光学遥感图像目标检测这一主题，面向标注信息不充足和标注信息不确切这两种标注信息受限情境，研究如何高效利用数据的标注信息，提升模型的检测性能。本文的创新性研究成果主要有：

(1) 针对现有边界框标注信息不足以引导模型精确检测的问题，提出了一种显著性信息引导的光学遥感图像舰船目标检测模型（Ship-S）。首先，创新性地设计了显著性预测分支引导的舰船目标检测联合优化算法，通过构建通道递减的映射结构和特征增强结构的方式来将显著性信息融入训练模型，提升了复杂背景下舰船目标表征的判别能力；针对训练正样本较少的问题，提出了显著性信息感知的样本扩充采样算法，综合考虑锚点位置的交并比（IoU）和显著性信息，将忽略集中的潜在正样本进行回收，实现了正例样本的有效扩增。构建了2个新的光学遥感图像数据集 HRSC-SO 和 DOTA-isaid-ship，实验结果证明 Ship-S能够取得优异性能表现，并具有较高的计算效率和较低的内存占用。

(2) 针对现有标注信息不充足所造成的舰船定位不准确和小型舰船漏检等问题，提出了一种标注信息增强利用的光学遥感图像舰船目标检测模型（SASOD）。首先，创新性地设计了尺度匹配和结构匹配的双匹配算法，通过分辨率匹配的多层监督学习，实现显著性标注信息的高效利用；其次，提出了一种跨阶段特征交互算法，通过通道级的连接，实现高层特征与低层特征的有效融合；最后，提出了一种交并比（IoU）自适应的标签分配算法，借助动态调整小目标的 IoU 分配阈值，实现了定位标注信息对舰船目标表征的有效增强。在光学遥感图像数据集HRSC-SO 和 DOTA-isaid-ship 上进行实验，实验结果证实 SASOD 对小型舰船的检测效果明显优于现有典型方法，在 DOTA-isaid-ship 上证实了模型中显著性分支的性能优势。

(3) 针对仅含有图像级类别标注的目标检测任务中存在的目标检测区域不完整以及漏检问题，提出了一种实例双优化的光学遥感图像弱监督目标检测模型（IDO）。首先，创新性地提出了一种实例挑选策略优化算法，通过课程学习的范式对带噪样本池进行递进式筛选，能够实现伪标签的去噪；其次，提出了一种实例特征全局优化算法，将类别激活图作为先验信息，对特征进行权重调整，实现低置信目标的全局显著信息补充。在 2 个主流的光学遥感数据集 NWPU VHR-10.v2 和 DIOR 上进行实验验证，结果表明所提算法在检测效果和定位性能两个指标上均具有良好的效果。另外还在 4 个大型目标检测数据集 PASCAL VOC 2007、 PASCAL VOC 2012、 MS COCO 2014 和 MS COCO 2017 上，证明了算法的泛化性。

Remote sensing technology is a non-contact, long-range detection technology that can acquire information in different wavelength bands and at different resolutions through sensors or remote sensors mounted on specific platforms. Among them, optical remote sensing technology occupies an important position. With the development of remote sensing platform technology, the optical remote sensing image data is growing explosively. How to efficiently and accurately decipher these data has become an urgent problem in remote sensing. In this context, object detection technology plays a crucial role as the cornerstone of remote sensing image interpretation. It can not only accurately identify the location of specific targets in the image, but also finely categorize these objects, which in turn provides solid data support for urban planning, environmental monitoring, resource management, and other applications. Therefore, it is of great theoretical and practical application value to carry out research on target detection in optical remote-sensing images.
In recent years, research on object detection in optical remote sensing images has undergone a significant transformation from traditional methods to deep learning approaches and has achieved breakthroughs in real-time and accuracy. However, with the increasing complexity and specialization of application scenarios, the issue of limited labels has gradually emerged as a key factor restricting the improvement of object detection performance. First, insufficient annotation information results in model performance that fails to meet the high-precision demands of specific application scenarios. Second, inexact annotation information leads to the model learning incorrect feature representations. Therefore, this dissertation focuses on the theme of object detection with limited labels in optical remote sensing images, addressing situations where annotation information is insufficient or inexact, and investigates how to efficiently utilize labels to enhance the detection performance of models. The innovative research outcomes of this dissertation include:

(1) Aiming at the problem that the existing label information of the bounding box is insufficient to guide the model for accurate detection, a saliency information-guided ship object detection model (Ship-S) in optical remote sensing images is proposed. First, a joint optimization algorithm for ship object detection guided by saliency prediction branches is innovatively designed to incorporate saliency information into the training model by constructing a channel-decreasing mapping structure and a feature-enhancing structure, which improves the discriminative ability of ship object characterization in a complex background. Additionally, to tackle the problem of fewer positive training samples, a saliency-aware sample augmentation sampling algorithm is proposed. By comprehensively considering the intersection over union (IoU) and saliency information of anchors, this algorithm recovers potential positive samples that have been ignored and effectively augments positive samples. Two new optical remote sensing image datasets, i.e., HRSC-SO and DOTA-isaid-ship, are constructed. Experimental results demonstrate that Ship-S achieves excellent performance and exhibits high computational efficiency and low memory usage.

(2) Aiming at the problems of inaccurate ship localization and missed detection of small ships caused by insufficient annotation information, a ship object detection model with enhanced utilization of annotation information in optical remote sensing images (SASOD) is proposed. First, a dual matching algorithm of scale matching and structure matching is innovatively designed to realize the efficient utilization of salient annotation information through multilayer supervised learning of resolution matching. Second, a cross-stage interaction algorithm is proposed to facilitate the effective fusion of high-level and low-level features through channel-level connections. Finally, an adaptive IoU-based label assignment algorithm is presented, dynamically adjusting the IoU threshold for small objects to enhance the impact of localization labels on ship representation. Experiments conducted on two optical remote sensing image datasets, i.e., HRSC-SO and DOTA-isaid-ship, confirm the superior detection performance of SASOD for small ships compared to existing methods. Additionally, the effectiveness of the saliency branch within the model is demonstrated on the DOTA-isaid-ship dataset.

(3) A weakly supervised object detection model with instance dual-optimization in optical remote sensing images (IDO) is proposed to address the problems of incomplete object detection region and missed detection in object detection tasks containing only image-level category labels. First, an instance selection strategy-optimization algorithm is innovatively proposed, which can achieve the denoising of pseudo-labels by incrementally screening the pool of noisy samples through the curriculum learning paradigm; second, an instance feature global optimization algorithm is proposed, which takes the class activation maps as the a priori information and adjusts the features with the weights, to achieve the supplementation of the globally significant information of the low-confidence targets. Experimental validation is carried out on 2 mainstream optical remote sensing datasets, NWPU VHR-10.v2 and DIOR, and the results show that the proposed algorithm has good results in both detection effect and localization performance. The generalizability of the algorithm is also demonstrated on four large object detection datasets PASCAL VOC 2007, PASCAL VOC 2012, MS COCO 2014, and MS COCO 2017.