摘要
静爆试验威力大、范围广,自动报靶系统无法布设在靶板周围实现穿孔识别,而人工识别方式效率低下,基于传统Otsu视觉算法的靶板穿孔识别精度不够,针对该问题,提出了一种基于图像配准和深度学习的靶板穿孔识别方法。采用双重图像配准方法提高试验前后靶板图像的配准精度,从差分图像中提取穿孔区域,并用ResNet50-NAM深度学习网络对穿孔区域进行分类识别。结果表明,所提方法的平均穿孔识别准确率为97.26%,与传统Otsu视觉算法相比,平均穿孔识别精度提高20.95%,平均穿孔识别时间比人工识别方式缩短83.89%,能够为武器装备毁伤效能评估提供快速、准确的数据支撑。
The static explosion test has great power with wide range.The automatic target-scoring system can not be installed close to the metal plate to realize penetrated hole detection,and the manual recognition method is inefficient,the traditional Otsu vision algorithm has a low detection accuracy for the penetrated hole.To solve those problems,a penetrated hole detection method for metal plate based on image registration and deep learning is proposed.The double image registration method is used to improve the registration accuracy of the metal plate image before and after the test,and the penetrated hole’s area is extracted from the differential image.The penetrated hole’s area is then classified and detected by using ResNet50-NAM deep learning network.The results show that the average penetrated hole detection accuracy of the proposed method is 97.26%and compared with the traditional Otsu vision algorithm,the average detection accuracy is increased by 20.95%.The average detection time is reduced by 83.89%compared with the manual recognition,which can provide rapid and accurate data support for evaluating the damage effectiveness of equipment.
作者
林威
李超
王晓坤
赵小强
张见升
LIN Wei;LI Chao;WANG Xiaokun;ZHAO Xiaoqiang;ZHANG Jiansheng(NORICO Group Test and Measuring Academy,Xi’an 710043,China)
出处
《兵器装备工程学报》
北大核心
2025年第7期296-302,共7页
Journal of Ordnance Equipment Engineering
基金
国防科技创新特区项目(20-163-30-ZT-004-015-01)。
关键词
穿孔识别
图像配准
NAM注意力
图像处理
penetrated hole detection
image registration
normalization-based attention module(NAM)
image processing
