摘要
线圈炮的管道内壁长期工作在高温、高压及强电磁作用下,容易形成多种烧蚀疵病而使线圈炮效能、安全及服役寿命降低,而人工检测缺陷效率低下。针对此问题,本文提出基于深度学习来实现线圈炮缺陷的自动检测与分类。首先采用双边滤波算法进行样本图像增强,降低噪声对成像的影响,并基于单样本几何变换的数据增强策略扩充样本数量,提高模型的适应性。然后比较了Alexnet、VGG16和Resnet18这三种经典深度卷积神经网络模型的性能,结果表明三个模型的平均类识别率分别为94.8%、94.5%和94.5%,单张图像缺陷检出率分别为94.1%、93.3%和94.7%,三者的结果相差不大。但在对480张测试集图像预测的花费总时间上,Alexnet仅需要0.32秒即可完成全部预测,远远少于VGG16和Resnet18所花费的时间,二者分别用时9.19秒和11.04秒。随后引入了Grad-CAM表征每个像素对该类图像的重要程度,结果显示模型对图像缺陷区域更敏感。最后基于C#语言开发了能自动化完成模型训练与缺陷分类的软件,大大提高了检测效率。
The inner wall of the coilgun’s pipeline is prone to form a variety of ablation defects under the action of high tem⁃perature,high pressure and strong electromagnetic for a long time,which reduces the efficiency,safety and service life of the coil⁃gun,but the efficiency of manual defect detection is low.Aiming at this problem,it is proposed to realize automatic detection and classification of coil gun defects based on deep learning.First,the bilateral filtering algorithm is used for image enhancement to re⁃duce the impact of noise on imaging,and the data enhancement strategy based on single-sample geometric transformation is used to expand the number of samples to improve the adaptability of the model.Then compare the performance of the three classic deep convolutional neural network models of Alexnet,VGG16 and Resnet18.The results show that the average class recognition rates of the three models are 94.8%,94.5%,and 94.5%,respectively,and the single image defect detection rates are 94.1%,93.3%and 94.7%,respectively.The results of the three are not much different.But in the total time for predicting 480 test set images,Alexnet only needs 0.32 seconds,which is far less than the time spent by VGG16 and Resnet18 is 9.19 seconds and 11.04 seconds respec⁃tively.And the introduction of Grad-CAM to characterize the importance of each pixel to this type of image,the results show that the model is more sensitive to image defect areas.Finally,based on the C#language,a software that can automatically complete the model training and defect classification is developed,which greatly improves the detection efficiency.
作者
田浩杰
杨晓庆
翟晓雨
Tian Haojie;Yang Xiaoqing;Zhai Xiaoyu(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065)
出处
《现代计算机》
2022年第10期86-91,共6页
Modern Computer
基金
武器装备技术基础科研项目(JZX7J20191201ZL0002)。
关键词
线圈炮
深度学习
缺陷检测
图像增强
coilgun
deep learning
defect detection
image enhancement
