近年来，基于深度学习的遥感影像目标检测取得了巨大进展，然而，遥感影像的标注工作需要大量的人力物力，且依赖于专家知识。遥感影像集精确标注的缺失给目标检测任务带来了巨大的挑战。基于上述问题，本文进行了以下研究工作：

（1） 基于遥感影像高分辨率、目标小而密集、背景复杂等特点，提出了基于跨特征融合与级联反馈的深度显著性目标检测网络，并基于像素位置信息优化了损失函数，有效地保留了结果图的细节特征，抑制了噪声与复杂背景的干扰；

（2） 引入基于域差异融合的自适应层结构，将自然图像的目标检测数据集应用到遥感网络训练过程中，并通过最大均值差异算法计算域适应损失，使网络更加适应遥感数据分布的特征，从而利用知识迁移解决遥感训练数据标注缺失的问题，进一步提高目标检测的准确度。

大量的对比实验与消融实验结果证明，本文算法能够保留小目标的特征信息，排除噪声与复杂背景的干扰，有效提取遥感影像目标信息。

Remote sensing images (RSIs) contain lots of effective information. In recent years, CNN-based object detection of RSIs have achieved great success. However, the annotation of RSIs requires huge time and cost. Lack of annotation brings great challenges to the task of RSI object detection. Based on the question above, the main research work of this paper can be summarized as follows:

(1) Based on the characteristic of RSIs, we propose a salienct object detection model based on multi-layer feature fusion, cascade feedback module and pixel-weighted loss function. The model preserve the detail of small and densely distributed targets, and suppress the complex background and noise.

(2) We introduce a domain adaptation module based on domain discrepancy fusion, which effectively transfers knowledge from natural domain to remote sensing domain and overcomes the lacking of annotations in remote sensing domain.

Plenty of comparative and ablation experiments have been applied in this paper. Results show that our algorithm performs better than state-of-the-arts.