毕业生知识库

请

高分辨率遥感图像变化类型识别是遥感图像处理领域的研究热点。该任务旨在识别和定位遥感图像中语义上的变化区域和变化类型。变化类型的识别在城市规划、森林监测、灾害评估等诸多领域具有丰富的应用场景和应用需求。

变化类型识别是一个具有挑战性和研究难点的计算机视觉任务。由于多时相遥感图像在获取的时候会受到光照、季节等因素影响，伪变化不可避免地会出现，导致未变化的区域像素上的差异也会非常明显。同时，由于地物类型特征存在类内差异大、类间差异小的现象，变化的类型判断也十分困难。目前的主流方法采用深度学习网络解决这一问题。但其本身具有很多局限性，包括网络模型地物特征提取能力不足、数据集自身缺陷导致训练不充分、缺乏与遥感图像处理其他任务的交互等问题。

针对上述难点，本论文做了两方面的工作。首先，变化类型识别任务的一个重要难点在于数据集中变化区域和非变化区域之间以及不同变化类型之间的数据不平衡。这一现象极大地影响了网络训练过程和最终性能。为了解决这个问题，本文提出了一种基于双网络结构和累积学习策略的新方法。在双网络结构的帮助下，深度学习网络在平衡非变化区域和变化区域方面更加稳健。通过累积学习策略，网络训练过程更加稳定。大量的实验证明了所提出的方法在各种变化检测数据集和现有的变化检测框架上的有效性。该方法可以显著提高提高模型进行变化类型识别的能力。

变化类型识别任务同时需要变化检测和语义分割的结果。变化检测任务和语义分割任务的原理、方法和应用密切相关。现有的方法忽略了任务间的联系，因此本文提出了一种新的方法来实现变化检测和语义分割的协同学习提高模型处理变化类型识别任务的能力。通过探索变化检测和语义分割之间的相关性和一致性，设计合理的结构保障了双任务间的信息交流和对模型的共同约束。该方法同步提高了网络的语义分割和变化检测能力，优于单一的变化检测网络或语义分割网络。具体而言，该方法通过共享的骨干网络提取多层级双时相特征，并通过两个独立的逐层解码器同时获得变化检测结果和语义分割结果。此外，设计了交互融合模块对变化信息和语义信息进行融合，计算了对比损失增强了两个任务之间的约束。通过在两类数据集上的主体实验和消融实验证明了该方法的有效性。

综上所述，本文通过双网络结构和累积学习策略缓解了数据不平衡的问题，提出了一种基于协同学习框架的变化类型识别模型。

输入中文摘要

The recognition of change types in high-resolution remote sensing images is a research hotspot in the field of remote sensing image processing. This task aims to identify and locate semantically changing regions and types in remote sensing images. The identification of change types has rich application scenarios and demands in various fields such as urban planning, forest monitoring, and disaster assessment.

This task is a challenging and difficult task. Due to the influence of factors such as illumination and season on the acquisition of remote sensing images, pseudo changes inevitably occur, resulting in significant differences in pixels even in unchange regions. At the same time, due to the phenomenon of large intra class differences and small inter class differences in the characteristics of surface features, it is also very difficult to determine the type of change. The current mainstream method uses deep learning networks to solve this problem. However, it has many limitations, including insufficient ability to extract features from network models, inadequate training due to the shortcomings of the dataset itself, and lack of interaction with other tasks in remote sensing image processing.

In response to the above difficulties, this paper has done two aspects of work. Firstly, an important challenge in the task of identifying change types lies in the data imbalance between changing and non changing regions in the dataset, as well as between different change types. This phenomenon greatly affects the network training process and final performance. To address this issue, this paper proposes a new method based on a dual network structure and cumulative learning strategy. With the help of a dual network structure, deep learning networks are more robust in balancing non changing and changing regions. By using cumulative learning strategies, the network training process becomes more stable. Numerous experiments have demonstrated the effectiveness of the proposed method on various change detection datasets and existing change detection frameworks. This method can significantly improve the model's ability to recognize change types.

The task of identifying change types requires both change detection and semantic segmentation results. At the same time, the principles, methods, and applications of change detection tasks and semantic segmentation tasks are closely related. The existing methods ignore the relationship between tasks, so this paper proposes a new method to implement the collaborative learning of change detection and semantic segmentation to improve the ability of the model to deal with change type recognition tasks. By exploring the correlation and consistency between change detection and semantic segmentation, a reasonable structure was designed to ensure information exchange and common constraints on the model between the two tasks. This method synchronously improves the semantic segmentation and change detection capabilities of the network, which is superior to a single change detection network or semantic segmentation network. Specifically, the method extracts multi-level and dual temporal features through a shared backbone network, and simultaneously obtains change detection results and semantic segmentation results through two independent layer by layer decoders. In addition, the interactive fusion module is designed to fuse the change information and Semantic information, and the contrast loss is calculated to enhance the constraints between the two tasks. The effectiveness of this method has been demonstrated through subject experiments and ablation experiments on two types of datasets.

In conclusion, this paper alleviates the problem of data imbalance through dual network structure and cumulative learning strategy, and proposes a change type recognition model based on collaborative learning framework.