Defect Detection of Spray Printed Variable Color 2D Code Based on ResNet34-TE

Fig. 5. Framework of the ResNet34-TE model. (a) Overall structure of the model; (b) structure of the residual module of model; (c) structure of Transformer-encoder

Download full size

Fig. 6. Process of position embedding

Download full size

Fig. 7. Structure of encoder. (a) Overall structure of encoder; (b) structure of multi-head attention mechanism; (c) structure of the multilayer perceptron block

Download full size

Fig. 8. Flowchart of proposed algorithm

Download full size

Fig. 9. Comparison of results of different models on validation set. (a) Loss curve; (b) accuracy curve

Download full size

Fig. 10. Confusion matrices of contrasting models. (a) Confusion matrix for ResNet34 model; (b) confusion matrix for ResNet34-TE model

Download full size

Fig. 11. Prediction results of various defects under different light intensities. (a) Flying ink defect; (b) ink smudging and dirt defect; (c) wire defect; (d) leakage defect

Download full size

Fig. 12. Visual heat maps of model before and after improvement

Download full size

Fig. 13. Sample images of six defects from the NEU-DET dataset

Download full size

Table 1. Dataset distribution
View table
Table 1. Dataset distribution
Defect type Enhanced training set Original training set Validation set Test set
Total 2378 1203 402 401
LX 620 311 102 99
FM 601 314 101 103
HM 593 304 100 101
LY 564 274 99 97

Table 2. Comparison of Classification Performance of Classical Models

View table

Table 2. Comparison of Classification Performance of Classical Models

Model	Accuracy /%	Precision /%	Recall /%	F1-score /%	Params /MB	CPU /ms
AlexNet	66.32	67.55	67.96	67.67	217.0	9.26
VGG16	85.00	85.98	85.03	85.94	532.0	35.08
MobileNet-V2	90.90	91.03	91.43	91.56	8.3	9.56
ShuffleNet-V2	79.88	79.94	80.68	79.67	6.5	10.69
ResNet34	91.52	91.96	92.02	91.96	81.3	16.56
ResNet50	93.30	94.36	92.58	92.20	90.0	24.84
EfficientNet-B0^［26］	95.26	95.51	94.97	95.18	16.8	17.82
MobileViT^［27］	96.37	96.78	96.40	96.65	10.8	21.85
ResNet34-TE	96.80	96.89	97.04	96.93	25.4	15.59

Table 3. Defect classification results of model before and after improvement
View table
Table 3. Defect classification results of model before and after improvement
Defect type ResNet34 ResNet34-TE
Accuracy /% Recall /% F1-score /% Accuracy /% Recall /% F1-score /%
FM 90 90 90 96 96 97
HM 87 84 86 98 94 95
LX 90 96 93 96 98 97
LY 100 98 99 100 100 100

Table 4. Different light test results
View table
Table 4. Different light test results
Defect type Light Accuracy /% Average accuracy /%
FM -30% 94.17 95.14
Standard 96.11
+30% 95.15
HM -30% 92.08 93.07
Standard 94.05
+30% 93.07
LX -30% 96.97 97.64
Standard 97.98
+30% 97.98
LY -30% 98.97 99.65
Standard 100.00
+30% 100.00

Table 5. Results of different encoder block depths
View table
Table 5. Results of different encoder block depths
Depths Accuracy /% Loss Params /MB
1 93.5 0.1054 22.1
2 94.3 0.0843 22.9
3 94.9 0.0734 23.7
4 96.1 0.0576 24.6
5 96.8 0.0428 25.4
6 94.7 0.0773 26.5
7 94.2 0.0815 27.2

Table 6. Results of different models on the NEU-DET dataset
View table
Table 6. Results of different models on the NEU-DET dataset
Model Accuracy /% Precision % Recall /% F1-score /%
ResNet34 96.52 96.92 96.61 96.62
ViT 92.14 92.78 91.86 92.54
CNN-PCA-DT^［30］ 98.51 97.41 97.82 98.34
TARGAN^［31］ 98.25 98.50 98.10 98.30
ResNet-TE 98.86 98.10 98.49 98.97

Tools

Get Citation

Copy Citation Text

Ying Li, Yao Dong, Zifen He, Hao Yuan, Fuyang Sun, Lingxi Gong. Defect Detection of Spray Printed Variable Color 2D Code Based on ResNet34-TE[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1812003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites