Automatic change detection from remotely sensed imagery is extremely important for many applications, including land use mapping. In recent years, a growing number of researchers have applied capable deep-learning methods to the research on change detection. The majority of deep learning-based change detection methods currently perform pixel-by-pixel classification at the original image scale, but they can hardly avoid the false changes caused by strong parallax effects and projected shadows, without considering the totality of changed objects/regions. In this study, we propose an object-level change detection framework to detect changed geographic entities (such as newly built buildings or changed artificial structures) by paying more attention to the overall characteristics and context association of changed object instances. The detected changed objects are represented as bounding boxes, which are simple, regular, and convenient to use in object feature extraction. In terms of data handling, a special data augmentation method for change detection called Alternative-Mosaic is proposed to effectively accelerate model training and improve model performance. For the model, we propose a one-stage change detection network called dual correlation attention-guided detector (DCA-Det) to detect the changed objects. In particular, we feed the dual-temporal images into a weight-shared backbone network to extract the change features of different scales. The change features on the same scale are further refined, and then the features between different scales are fused by the correlation attention-guided feature fusion neck. Finally, the change detection heads output the prediction results of the changed objects/regions of different scales. Experiments were conducted on public LEVIR building change detection and aerial imagery change detection (AICD) datasets. The quantitative evaluation and visualization results proved the superiority and robustness of our framework. Our DCA-Det can obtain state-of-the-art performance on object-level metrics (99.50% ${\mathit{AP}}_{\mathit{IoU}=.50}$ and 79.72% ${\mathit{AP}}_{\mathit{IoU}=.50:.05:.95}$) on the AICD-2012 dataset.

对于包括土地用途制图在内的许多应用，从遥感影像中自动检测变化非常重要。近年来，越来越多的研究人员将功能强大的深度学习方法应用于变更检测的研究。目前，大多数基于深度学习的变化检测方法都在原始图像尺度上执行逐像素分类，但是，如果不考虑变化对象/区域的总数，它们几乎无法避免由于强烈的视差效应和投影阴影而导致的虚假变化。 。在这项研究中，我们提出了一个对象级别的变化检测框架，通过更加关注变化对象实例的整体特征和上下文关联，来检测变化的地理实体（例如新建建筑物或变化的人工结构）。检测到的已更改对象表示为边界框，边界框简单，规则且易于在对象特征提取中使用。在数据处理方面，提出了一种用于变更检测的特殊数据增强方法，称为“ Alternative-Mosaic”，以有效地加速模型训练并提高模型性能。对于该模型，我们提出了一种称为双相关注意力引导检测器（DCA-Det）的单阶段变化检测网络，用于检测变化的对象。特别是，我们将双时相图像输入到一个权重共享的主干网络中，以提取不同尺度的变化特征。进一步细化相同尺度上的变化特征，然后通过相关性关注引导特征融合颈部融合不同尺度之间的特征。最后，变化检测头输出不同比例的变化对象/区域的预测结果。对公共LEVIR建筑物变更检测和航空影像变更检测（AICD）数据集进行了实验。定量评估和可视化结果证明了我们框架的优越性和鲁棒性。我们的DCA-Det可以在对象级指标上获得最先进的性能（99.50％${\mathit{美联社}}_{\mathit{o}=。50}$ 和79.72％ ${\mathit{美联社}}_{\mathit{o}=。50：。05：。95}$）放在AICD-2012数据集上。