基于改进Faster-RCNN的机场场面小目标物体检测算法

针对目前应用于机场视频监控中的卷积神经网络方法存在小目标物体识别准确率较低的问题,本文提出了一种基于Faster-RCNN并结合多尺度特征融合与在线难例挖掘的机场场面小目标检测算法。该算法采用ResNet-101作为特征提取网络,并在该网络基础上建立了一个带有上采样的"自顶向下"的特征融合模块,以生成语义信息更加丰富的高分辨率特征图。并在网络训练过程中,采用在线难例挖掘的训练策略使模型更加鲁棒地对小目标样本进行定位。最后,手工构建了一个包含5 982张图片的机场数据集,用于检测模型的训练和测试。结果表明,本文所提出算法显著提升了机场场面小目标物体检测的准确率,且使整体平均检测准确率达到了80.8%,该结果高于其他先进的目标检测模型。

Small Target Detection in Airport Scene via Modified Faster-RCNN

For the low precision of small target detection in the existing convolution neural network methods used in airport video surveillance, in this paper, a small target detection algorithm based on Faster-RCNN combined with multi-scale feature fusion and online-hard-example-mining(OHEM) is proposed. First of all, ResNet-101 is adopted as the feature extraction backbone, and a top-down multi-scale feature fusion pathway is established based on the ResNet-101 to generate richer semantic feature maps of a fine resolution. During the network training, OHEM is adopted to make the network more robust to locate the region of small target objects. At last, an airport dataset containing 5 982 pictures is constructed manually, which is used to verify the training and testing of the model. The results show that our modified Faster-RCNN algorithm significantly improves the accuracy of small target detection under airport situation. Besides, the mean average precision reaches 80.8%, which is higher than other advanced object detection models.