基于卷积神经网络的短切毡缺陷分类 下载: 1008次
Classification of Chopped Strand Mat Defects Based on Convolutional Neural Network
西安工程大学电子信息学院, 陕西 西安 710048
图 & 表
图 1. 卷积神经网络模型结构
Fig. 1. Architecture of convolutional neural network model
下载图片 查看原文
图 2. 卷积神经网络训练过程
Fig. 2. Training process of convolutional neural network
下载图片 查看原文
图 3. 实验流程
Fig. 3. Flow chart of experiment
下载图片 查看原文
图 4. 缺陷样本数量
Fig. 4. Number of defect samples
下载图片 查看原文
图 5. 数据集样本。(a)~(d)并条;(e)~(h)分散不良;(i)~(l)纱结;(m)~(p)污渍
Fig. 5. Dataset samples. (a)-(d) Parallel; (e)-(h) poor dispersion; (i)-(l) yarn knot; (m)-(p) stain
下载图片 查看原文
图 6. 参数初始化方法对比。(a)训练精度;(b)验证精度;(c)训练损失值;(d)验证损失值
Fig. 6. Comparison of initialization parameters. (a) Training accuracy; (b) verification accuracy; (c) training loss values; (d) verification loss values
下载图片 查看原文
图 7. 网络微调训练过程。(a)训练精度;(b)验证精度;(c)训练损失值;(d)验证损失值
Fig. 7. Training process of fine-tuning. (a) Training accuracy; (b) verification accuracy; (c) training loss values; (d) verification loss values
下载图片 查看原文
表 1参数初始化方法对比
Table1. Comparison of initialization parameters
Method | Trainingaccuracy /% | Trainingloss | Verificationaccuracy /% | Verificationloss |
---|
Randomlyinitialized | 80.0 | 0.0270 | 75.0 | 0.032 | Migrationinitialized | 99.6 | 0.0016 | 86.5 | 0.023 |
|
查看原文
表 2网络模型性能
Table2. Model performances
Network structure | Training accuracy /% | Training loss | Verification accuracy /% | Verification loss | Modeling time /s |
---|
Resnet18 | 99.7 | 0.0019 | 84.6 | 0.026 | 350 | Resnet50 | 99.9 | 0.0010 | 93.0 | 0.015 | 949 | Resnet101 | 99.8 | 0.0058 | 85.6 | 0.022 | 1518 | VGG11 | 99.2 | 0.0017 | 80.0 | 0.035 | 940 | VGG16 | 99.4 | 0.0013 | 86.2 | 0.027 | 1629 | VGG19 | 99.6 | 0.0016 | 86.5 | 0.023 | 1952 |
|
查看原文
表 3测试样本混淆矩阵
Table3. Confusion matrix of test sample
Matrix | Predicted value |
---|
Parallel | Poordispersion | Yarnknot | Stain |
---|
Truevalue | Parallel | 260 | 2 | 0 | 0 | Poordispersion | 0 | 257 | 15 | 0 | Yarn knot | 0 | 1 | 243 | 0 | Stain | 0 | 0 | 2 | 260 |
|
查看原文
表 4网络性能评价指标
Table4. Evaluation of network performances
Defect category | P | R | |
---|
Parallel | 1.00 | 0.99 | 1.00 | Poor dispersion | 0.99 | 0.94 | 0.97 | Yarn knot | 0.93 | 1.00 | 0.96 | Stain | 1.00 | 0.99 | 1.00 |
|
查看原文
卓东, 景军锋, 张缓缓, 苏泽斌. 基于卷积神经网络的短切毡缺陷分类[J]. 激光与光电子学进展, 2019, 56(10): 101009. Dong Zhuo, Junfeng Jing, Huanhuan Zhang, Zebin Su. Classification of Chopped Strand Mat Defects Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(10): 101009.