针对现有遥感图像目标检测算法对于复杂场景下多尺度目标检测精度较低、泛化能力差的问题,提出了一种多尺度卷积神经网络遥感目标检测框架——MSCNN。该方法通过构造一种深度特征金字塔,增强了网络对多尺度目标特征的提取能力;引入聚焦分类损失作为分类损失函数,加强了网络对难样本的学习能力。所提方法在NWPU VHR-10公开数据集上取得了0.960的平均检测精度(mAP),相较于RetinaNet检测框架,MSCNN对小尺度以及中等尺度目标的平均检测精度分别提高了1.5%和1.9%,实现了对多尺度目标的高精度稳健检测。

A novel detection framework is proposed based on a multiscale convolutional neural network (MSCNN) to overcome low precision and insufficient generalization ability associated with existing object detection methods for multiscale objects with complex scenes. First, an essence feature pyramid network is constructed to enhance the extraction ability of multiscale features. Then, the focal classification loss is introduced as classification loss function to enhance the learning capability of the MSCNN over complex samples. The proposed method achieves a mean average precision(mAP) of 0.960 over the challenging NWPU VHR-10 dataset. In comparison with the RetinaNet detection method, the mAP of the proposed MSCNN on small- and medium-scale objects increases by 1.5% and 1.9%, respectively. The proposed method is found to be accurate and robust for multiscale objects.