학술논문
基于深度学习的炸点图像识别与处理方法 / A Deep Learning Based Explosion Point Image Recognition and Processing Method
Document Type
Academic Journal
Author
Source
探测与控制学报 / Journal of Detection & Control. 46(1):70-77
Subject
Language
Chinese
ISSN
1008-1194
Abstract
为了提高弹丸近炸落点位置的测试精度,改善传统炸点图像易受噪声干扰及环境影响的识别问题,提出基于深度学习的炸点图像识别与处理方法.该方法是以高速摄像机拍摄的多序列炸点图像为基础,利用GoogLeNet分类网络方法提取爆炸瞬间炸点信息,研究改进 U-Net网络分割炸点图像,重点对炸点图像的主干特征提取和优化损失函数进行建模,对炸点图像数据信息集进行训练与测试,并通过Canny边缘提取算法提取炸点图像边缘,采用最小二乘法进行轮廓拟合,求解炸点像素坐标,结合摄像机空间几何关系,获得炸点空间坐标.实验结果表明,改进 U-Net 网络的 PA 值为 94.2%,MPA 值为 97.6%,MIOU 值为 84.8%,相比于原始U-Net网络的分割精度更高,能够为后续炸点位置的获取及武器毁伤评估提供技术支撑.
To improve the testing accuracy of the projectile's near-impact point and overcome the recognition deficien-cies caused by noise interference and environmental factors in traditional explosion point images,a deep learning-based explosion point image recognition and processing method was proposed in this paper.Based on multi-sequence explosion point images captured by a high-speed camera,the method utilized GoogLeNet classification network to extract explo-sion point information at the moment of explosion,the improvement of the U-Net network for segmenting explosion point images were studied,which focused on modeling the main feature extraction and optimization loss function of ex-plosion point images,then explosion point image datasets were trained and tested.The Canny edge detection algorithm was adopted to extract the edge of explosion point images,the least squares method was applied for contour fitting,and pixel coordinates of explosion points were solved.By combining camera's spatial geometry,spatial coordinates of the ex-plosion point were obtained.Experimental results demonstrated that the improved U-Net network achieved a segmenta-tion precision with a PA value of 94.2%,an MPA value of 97.6%,and an MIOU value of 84.8%,which was higher than that of original U-Net network.The proposed method could provide technical support for obtaining the location of explosion points.
To improve the testing accuracy of the projectile's near-impact point and overcome the recognition deficien-cies caused by noise interference and environmental factors in traditional explosion point images,a deep learning-based explosion point image recognition and processing method was proposed in this paper.Based on multi-sequence explosion point images captured by a high-speed camera,the method utilized GoogLeNet classification network to extract explo-sion point information at the moment of explosion,the improvement of the U-Net network for segmenting explosion point images were studied,which focused on modeling the main feature extraction and optimization loss function of ex-plosion point images,then explosion point image datasets were trained and tested.The Canny edge detection algorithm was adopted to extract the edge of explosion point images,the least squares method was applied for contour fitting,and pixel coordinates of explosion points were solved.By combining camera's spatial geometry,spatial coordinates of the ex-plosion point were obtained.Experimental results demonstrated that the improved U-Net network achieved a segmenta-tion precision with a PA value of 94.2%,an MPA value of 97.6%,and an MIOU value of 84.8%,which was higher than that of original U-Net network.The proposed method could provide technical support for obtaining the location of explosion points.