基于改进YOLO V3的海上弹着点水柱信号检测算法

Marine impact water column signal detection algorithm based on improved YOLO V3

在海上实弹射击训练的过程中,快速、准确地检测出弹着点处水柱信号对评估射击效果具有重要意义。针对传统的人工检靶方式效率低、误差大,且无法分辨重叠弹着点的问题,结合深度学习理论,提出了一种基于改进YOLO V3的目标检测算法。在网络输入端利用Mosaic数据增强方式,丰富了检测物体的背景和小目标,降低了网络训练门槛。将Mish函数作为网络基本组件中的激活函数,提高了网络泛化能力。构建新的检测模块,将输入的特征信息分为两个分支,通过特征压缩与拼接,实现通道间信息交互。同时,对不同尺度的特征进行融合,提高网络特征提取能力。目标数据集的实验结果表明,改进后的YOLO V3算法平均准确率提高了5.39%,达到了82.64%,检测速度由27.74 FPS提高到了29.61 FPS,可以更好地完成海上弹着点水柱信号检测任务。

During a live-fire training exercise at sea,it is very important to quickly and accurately detect the water column signal at the impact point to evaluate the firing effect.Aiming at the problems of low efficiency,large error and inability to distinguish overlapping impact points in the traditional manual target detection,a target detection algorithm based on improved YOLO V3 is proposed by combining with deep learning theory.Mosaic data enhancement method is used in the network input end to enrich the background and small targets of detected objects and reduce the threshold of network training.Mish function is used as activation function in basic network components to improve network generalization.A new detection module is constructed,which divides the input feature information into two branches to realize the information interaction between channels by compressing and splicing the features.At the same time,the different scales of features are fused to improve the ability of network feature extraction.Experimental results on the target data set show that the average accuracy of the improved YOLO V3 algorithm is improved by 5.39%,reaching 82.64%,and the detection speed is increased from 27.74 FPS to 29.61 FPS,which can better meet the detection requirements of water column signals at the impact point on the sea.