Optics and Precision Engineering, Volume. 31, Issue 22, 3357(2023)
A lightweight deep learning model for TFT-LCD circuits defect classification based on swin transformer
Figures & Tables(15)
Fig. 5. Conventional convolution and depthwise separable convolution
Fig. 8. Different effect of the parameters
Fig. 9. Different effect of the parameters
Fig. 10. Comparison of detection performance between ResNet-34 model and the improved model
Table 1. Results of parameter
n affects experimentsView tableView in ArticleTable 1. Results of parameter
n affects experimentsn Top-1 Acc FLOPs/G i/（m·s-1） 0 94.0 17.5 51.42 1 92.1 16.3 54.49 2 90.6 14.7 61.92
Table 2. Results of ablation experiments
View tableView in ArticleTable 2. Results of ablation experiments
Model Top-1 Acc/% Precision Recall F1 FLOPs/G i/（m·s-1） Swin-T 94.0 0.935 4 0.935 6 0.935 5 17.5 51.42 +Token merging 90.6 0.900 2 0.901 8 0.901 0 14.7 61.92 +DW 91.2 0.913 8 0.891 3 0.902 4 14.9 60.43 +KD 92.7 0.916 5 0.905 6 0.911 0 14.9 60.43
Table 3. Results of comparison experiments
View tableView in ArticleTable 3. Results of comparison experiments
Model Top-1 Acc/% FLOPs/G i/（m·s-1） Resnet-34 87.4 14.6 68.97 DenseNet169 91.1 13.7 32.72 EffecientNet-v2 90.6 11.6 28.57 MobileViT-v2 90.5 9.8 73.95 ConvNext-T 95.1 17.8 50.36 EVA02 95.8 87.5 10.62 Ours 92.7 14.9 60.43
Table 4. Results of comparison experiments on public dataset
View tableView in ArticleTable 4. Results of comparison experiments on public dataset
Model Size Top-1 Acc/% Params/M FLOPs/G EVA-G/14 336 89.5 1013.01 445.56 ViT-H/14 336 88.6 632.46 363.64 ConvNext-XL 384 87.7 350.20 179.03 Swin-L 384 87.1 196.74 100.28 ResNet-101d 320 83.0 44.57 23.82 MobileViT-v2 384 82.9 14.25 12.35 EffecientNet-v2 288 82.2 13.65 8.16 Ours 384 83.3 27.50 13.89
Tools
Get Citation
Copy Citation Text
Yan XIA, Chen LUO, Yijun ZHOU, Lei JIA. A lightweight deep learning model for TFT-LCD circuits defect classification based on swin transformer[J]. Optics and Precision Engineering, 2023, 31(22): 3357
Paper Information
Category:
Received: Mar. 30, 2023
Accepted: --
Published Online: Dec. 29, 2023
The Author Email: LUO Chen (chenluo@seu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20