基于轻量级网络的装甲目标快速检测

Fast Armored Target Detection Based on Lightweight Network

针对战场环境下装甲目标的检测任务,提出一种基于轻量级网络的快速检测方法.首先以轻量级卷积神经网络MobileNet 作为骨架网络,构建一个多尺度的单步检测网络;然后针对装甲目标的尺寸分布情况使用分辨率更高的卷积特征图,并在每个检测单元上新加入一个残差模块,增强了对小尺度目标的检测能力;最后引入focal-loss损失来替代传统的交叉熵损失函数,有效地克服了训练过程中存在的正负样本分布极度不平衡的问题.针对装甲目标构建了专用的目标检测数据集,并在该数据集上对几种目前主流的单步检测方法进行了训练和测试,实验结果表明,该方法在检测精度、模型容量以及运行速度上均取得了较好的效果,对于无人机等小型移动侦查平台具备良好的适用性.

Focused on the detection task of armored target in battlefield environment, a fast detection method based on lightweight convolutional neural network is proposed in this paper. Firstly, based on the lightweight backbone network (MobileNet), a multi-scale single-stage detection framework is developed. Secondly, considering the size distribution of armored target, higher resolution feature maps are selected and a new designed Resblock is added to each detection unit to enhance the detection performance for small targets. At last, focal-loss function is introduced to replace the traditional cross entropy loss function, which effectively overcomes the extreme imbalance of the distribution of the positive and negative samples in training processes. A special detection dataset for armored target is constructed, based on which the comparable experiments with state-of-art detection methods are conducted. Experimental results show that the proposed method achieves good performance in detection accuracy, model size and operation speed, and is especially suitable for small mobile reconnaissance platforms such as UAVs (unmanned aerial vehicle).