基于跳跃特征金字塔的域适应目标检测模型

Skip Feature Pyramid Based Domain Adapted Model for Object Detection

针对训练数据和测试数据来源不同,特征分布差异较大,以及目标大小不一时,深度目标检测模型性能下降明显的问题,提出基于跳跃特征金字塔的域适应遥感图像目标检测模型。首先,在目标检测网络中增加域适应部分,将训练数据作为源域,测试数据作为目标域,通过对抗训练的方式,对具有不同特征分布的两个域,关注其差异,同时提取源域和目标域数据的特征,减小遥感图像由光照、角度等不同造成的源域和目标域图像在图像级的域偏移对目标检测的影响,提升检测模型的推广性能。其次,提出跳跃特征金字塔结构,通过特征上采样,以及同层连接、相隔层间的跳跃连接与特征融合,来增强特征图的细节信息和语义信息,以提高模型对不同尺度目标的检测精度。最后,使用区域推荐网络在多个不同分辨率的特征图上提取候选区域,检测不同尺度的目标。所提模型在NWPUVHR-10数据集上检测精度达到了98.2%、误检率为5.4%、漏检率为8.3%;在RSOD-DATA的低亮度数据集上检测精度达到了62%,误检率、漏检率为18.2%和18.5%。与其他模型相比,所提模型的性能有明显提升,具有更好的推广性。

In order to solve the problem that the detection performance of the deep object detection model drops significantly when the feature distribution differ greatly between training data and testing data,and when objects vary in size,a skip feature pyramid based domain adapted model for remote sensing image object detection is proposed.Firstly,the domain adaptation part is added into the object detection network.Regarding training data as the source domain and testing data as the target domain and focusing on their differences,the network can extract robust deep convolution features for both source and target domains with different feature distributions,through the adversarial training.This reduces the influence of the deviation between source and target domain image due to different illumination and angle in image acquisition,and improves the generalization of the model.Secondly,a skip feature pyramid structure is proposed to enhance the detail information and semantic information of the feature map by feature upsample,connection and jump connection between the same layer and the nonadjacent layer,and feature fusion,so as to improve the detection accuracy for different scale objects.Finally,the region proposal network is used to extract candidate regions on multiple feature maps with different resolutions to detect targets of different scales.The experimental results show that the detection accuracy on the NWPUVHR-10 dataset is 98.2%,the false detection rate is 5.4%,and the missed detection rate is 8.3%.The accuracy on the RSOD-DATA low-brightness dataset reaches 62%,and the false detection rate and missed detection rate are 18.2%and 18.5%respectively.Compared with the other models,the performance of the proposed model has improved significantly with better generalization.