用于遥感图像目标检测的少样本算法

Few-shot Algorithm for Object Detection in Remote Sensing Images

针对遥感场景数据量匮乏,高空拍摄捕捉的地表物体尺寸变化明显,包含大量多个类别的物体以及复杂背景,导致检测准确率低、分类不准确等问题,提出一种基于二阶段检测模型(Faster RCNN)的少样本遥感目标检测网络。新增新型反转卷积算子构建检测器主干,提高特征提取能力;融入多尺度对象级正样本特征进行原始特征增强,抑制负样本的不利影响,充分挖掘各目标尺度的特征信息,帮助语义信息进行定位;采用对比监督的思想改进损失函数,细化目标分类,降低误检率。在公开遥感数据集上的实验结果表明,在仅有少量遥感标注样本的条件下,该网络能适应遥感图像的多尺度特征并有效缓解数据稀缺引发的过拟合现象。与先期Meta RCNN和FsDet网络相比,平均准确度进一步提升了3.8个百分点和2.5个百分点,为遥感领域的图像目标检测提供有意义参考。

In view of the lack of remote sensing scene data,the obvious size change of surface objects captured by aerial photography,including a large number of objects of multiple categories and complex background,resulting in low detection accuracy and inaccurate classification,a small sample remote sensing target detection network based on the two-stage detection model(Faster RCNN)is proposed.New involution convolution operators are added to build detector backbone to improve feature extraction capability;Integrate multi-scale object-level positive sample features to enhance the original features,suppress the adverse effects of negative samples,fully mine the feature information of each target scale,and help the semantic information to locate;The idea of comparative supervision is adopted to improve the loss function,refine the target classification and reduce the false detection rate.The experimental results on public remote sensing data sets show that the network can adapt to the multi-scale characteristics of remote sensing images and effectively alleviate the over-fitting phenomenon caused by data scarcity under the condition of only a small number of remote sensing labeled samples.Compared with the previous Meta RCNN and FsDet networks,the average accuracy has been further improved by 3.8 percentage points and 2.5 percentage points,providing a meaningful reference for image target detection in the remote sensing field.