全卷积网络多层特征融合的飞机快速检测

Fast Airplane Detection Based on Multi-Layer Feature Fusion of Fully Convolutional Networks

针对传统飞机检测方法准确率低、虚警率高、速度慢等问题,提出一种全卷积神经网络多层特征融合的飞机快速检测方法。将浅层和深层的特征经过采样后在同一尺度进行融合,以缓解由于深层特征图维度过低造成的对小目标表达不足的问题;修改区域提取时的选框尺寸以适应实际图像中飞机的尺寸特征;用卷积层代替全连接层以减少网络参数并适应不同大小的输入图像;复用区域提取网络和检测网络的卷积层和学习的特征参数以保证检测的高效性。仿真结果表明,与典型的飞机检测方法相比,所提方法在测试集上取得了更高的准确率和更低的虚警率,同时大大加快了检测速度。

In order to solve the problems of traditional airplane detection methods, such as low accuracy, high false alarm rate, and low speed, we propose a fast airplane detection method based on multi-layer feature fusion in a fully convolutional neural network. Firstly, we sample the shallow and deep features separately and fuse them at the same scale, which can alleviate the problem that the deep features are too sparse to express the small-size objects. Secondly, we redesign the size of the reference boxes to adjust to the practical size of the airplane in the input image. Thirdly, we replace the fully connected layers by convolutional layers to reduce the network parameters and adapt to input images with different sizes. Fourthly, we multiplex the convolutional layers and the learning-feature parameters of the proposal network and the detection network to improve the detection efficiency. The simulation results show that compared with typical airplane detection methods, the proposed method achieves higher accuracy and lower false alarm rate and greatly accelerates the detection speed.