采用反卷积与捷径连接,针对毫米波图像提出了一种高效、快速的卷积神经网络,在保留图像低阶细粒度特征的同时,检测速度由原框架的9 frame/s大幅提升至27 frame/s,并取消了Faster RCNN (Regions with Convolutional Neural Networks)中的RCNN部分。为了使网络更好地收敛,基于聚类思想设计了初始候选框的大小。使用在线困难样本挖掘(OHEM)优化了Faster RCNN的损失函数,解决了毫米波图像中正负样本失衡的问题,大幅提升了训练速度。所提算法在测试集上取得了87.6%的准确率和81.2%的检出率,F₁分数相较于主流算法提升了5%左右。

An efficient and fast convolution neural network for millimeter-wave images that uses deconvolution and a shortcut connection is proposed. The proposed network retains the low-order fine-grained features of the image and significantly improves the detection speed to 27 frame/s from 9 frame/s of original frame. The RCNN (Regions with Convolutional Neural Networks) part of the Faster RCNN is removed. To achieve better network convergence, the initial candidate box size is designed based on thought clustering. The online hard example mining is applied to optimize the loss function of the Faster RCNN such that the imbalance problem between positive and negative samples in millimeter wave images is solved and the training speed is improved significantly. By using the proposed algorithm, the accuracy of 87.6% and the detection rate of 81.2% are obtained on the test set. Compared with mainstream algorithms, the proposed algorithm improves the F1 score by approximately 5%.