Textile defect recognition network based on label embedding

Ying LIU; Wei JIANG; Guandian LI; Lei CHEN; Shuang ZHAO

doi:10.37188/OPE.20233110.1563

Journals >Optics and Precision Engineering >Volume 31 >Issue 10 >Page 1563 > Article

Optics and Precision Engineering
Vol. 31, Issue 10, 1563 (2023)

Textile defect recognition network based on label embedding

Ying LIU^*, Wei JIANG, Guandian LI, Lei CHEN, and Shuang ZHAO

Author Affiliations

College of Electronic Information Engineering， Changchun University of Science and Technology， Changchun130000， China

show less

DOI: 10.37188/OPE.20233110.1563 Cite this Article

Ying LIU, Wei JIANG, Guandian LI, Lei CHEN, Shuang ZHAO. Textile defect recognition network based on label embedding[J]. Optics and Precision Engineering, 2023, 31(10): 1563 Copy Citation Text

show less

Fig. 1. Structure of residual unit

Download full size | View in the Article

Fig. 2. Overall of TDRNet architecture when training

Download full size | View in the Article

Fig. 3. Overall of TDRNet Architecture When Testing

Download full size | View in the Article

Fig. 4. Example of textile defect category

Download full size | View in the Article

Fig. 5. Structure of Label Embedded Module

Download full size | View in the Article

Fig. 6. principle of DP loss

Download full size | View in the Article

Fig. 7. the dataset label distribution of Guangdong intelligent manufacturing category dataset

Download full size | View in the Article

Fig. 8. visualization of searching initial learning rate of TDRNet based on improved Resnet-50

Download full size | View in the Article

Fig. 9. visualization of learning rate schedule（γ=0.2） of TDRNet based on improved Resnet-50

Download full size | View in the Article

Fig. 10. loss curve of TDRNet based on improved Resnet-50

Download full size | View in the Article

Fig. 11. accuracy curve of TDRNet based on improved Resnet-50

Download full size | View in the Article

Layer name	Output size	Stride	Dilated rate	50-layer	101-laer
Input Layer	720×720	--	--	--
Conv0	360×360	2	1	$[\begin{matrix} 3 \times 3,64 \\ 3 \times 3,64 \\ 3 \times 3,128 \end{matrix}]$
Conv1_x	180×180	2	1	$3 \times 3, m a x p o o l$
Conv1_x	180×180	2	1	$[\begin{matrix} 1 \times 1,64 \\ 3 \times 3,64 \\ 1 \times 1,256 \end{matrix}] \times 3$	$[\begin{matrix} 1 \times 1,64 \\ 3 \times 3,64 \\ 1 \times 1,256 \end{matrix}] \times 3$
Conv2_x	90×90	2	2	$[\begin{matrix} 1 \times 1,128 \\ 3 \times 3,128 \\ 1 \times 1,512 \end{matrix}] \times 4$	$[\begin{matrix} 1 \times 1,128 \\ 3 \times 3,128 \\ 1 \times 1,512 \end{matrix}] \times 4$
Conv3_x	45×45	2	2	$[\begin{matrix} 1 \times 1,256 \\ 3 \times 3,256 \\ 1 \times 1,1 024 \end{matrix}] \times 6$	$[\begin{matrix} 1 \times 1,256 \\ 3 \times 3,256 \\ 1 \times 1,1 024 \end{matrix}] \times 23$
Conv4_x	23×23	2	2	$[\begin{matrix} 1 \times 1,512 \\ 3 \times 3,512 \\ 1 \times 1,2 048 \end{matrix}] \times 3$	$[\begin{matrix} 1 \times 1,512 \\ 3 \times 3,512 \\ 1 \times 1,2 048 \end{matrix}] \times 3$
Conv5	12×12	2	2	$[\begin{matrix} 1 \times 1,512 \\ 3 \times 3,512 \\ 1 \times 1,2 048 \end{matrix}] \times 1$	$[\begin{matrix} 1 \times 1,512 \\ 3 \times 3,512 \\ 1 \times 1,2 048 \end{matrix}] \times 1$
--	1×1	--	--	Global Average Pooling， FC，softmax

Table 1. Backbone of TDRNet

View in the Article

瑕疵名称	粗分类标签	细分类标签	瑕疵名称	粗分类标签	细分类标签
无疵点	0	0	星跳	16	18
破洞	1	1	跳花	16	19
水渍	2	2	断氨纶	17	20
油渍	2	3	稀密档	18	21
污渍	2	4	浪纹档	18	22
三丝	3	5	色差档	18	23
结头	4	6	磨痕	19	24
花板跳	5	7	轧痕	19	25
百脚	6	8	修痕	19	26
毛粒	7	9	烧毛痕	19	27
粗经	8	10	死皱	20	28
松经	9	11	云织	20	29
断经	10	12	双纬	20	30
吊经	11	13	双经	20	31
粗维	12	14	跳纱	20	32
纬缩	13	15	筘路	20	33
浆斑	14	16	纬纱不良	20	34
整经结	15	17	纬纱不良	20	34

Table 2. Defect Category of GDIM-CD

View in the Article

软硬件名称	硬件型号/软件版本
CPU	Intel（R） Core（R） i7-10700KF@3.8GHz
GPU	NVIDIA RTX3090
内存	Fury HX432C16FB3K2/32
操作系统	Ubuntu 20.04.3 LTS （GNU/Linux 5.11.0-37-generic x86_64）
CUDA版本	11.1.1
Python版本	3.9.6
Pytorch版本	1.9.0

Table 3. Experimental environment of TDRNet

View in the Article

Backbone	LEM Model	DP Loss	Seesaw Loss	Top1 err. （%）
ResNet-50 （baseline）	×	×	×	20.52
ResNet-50	√	×	×	20.59
ResNet-50	√	√	×	18.10
ResNet-50	×	×	√	18.97
ResNet-50	√	√	√	17.32
Improved ResNet-50	×	×	×	19.94
Improved ResNet-50 （TDRNet-50）	√	√	√	16.80
ResNet-101	×	×	×	19.58
ResNet-101	√	√	√	16.61
Improved ResNet-101	×	×	×	19.23
Improved ResNet-101 （TDRNet-101）	√	√	√	16.35

Table 4. Ablation study of TDRNet on rough-grained task

View in the Article

Model	Top1 err./%	Top5 err./%	Params./M	FLOPs/G	FPS
EfficientNet_B0^［25］	25.11	12.64	4.03	4.40	33
DenseNet-169^［26］	23.11	12.44	12.52	35.55	13
EfficientNet_B4^［25］	20.20	8.15	17.59	16.70	16
DenseNet-201^［26］	19.49	8.46	18.13	45.44	11
ResNext-50^［27］	20.13	7.34	23.02	44.74	26
ResNet-50^［18］	21.27	7.60	23.55	43.13	36
TDRNet-50	17.45	5.20	32.34	58.91	29
EfficientNet_B6^［25］	18.35	7.08	40.78	36.91	11
ResNet-101^［18］	20.04	7.79	42.54	81.75	21
TDRNet-101	17.12	5.27	51.33	97.53	19
AlexNet^［17］	27.34	11.18	57.09	7.37	277
WRN50^［28］	20.68	8.99	66.88	119.42	19
ViT_B_16^［29］	20.06	9.76	87.22	249.12	9
ViT_B_32^［29］	20.30	10.08	87.84	51.49	44
WRN101^［28］	19.13	8.63	124.88	237.08	11
VGG16^［30］	23.85	10.73	134.35	158.68	21
VGG19^［30］	24.01	10.54	139.66	201.67	18
ViT_L_16^［29］	21.55	11.09	305.20	815.69	3
ViT_L_32^［29］	21.39	11.38	306.02	175.66	17

Table 5. Comparison of the typical classification model experimental results for fine-grained task on GDIM-CD

View in the Article

Model	TDRNet	MA-CNN^［31］	RA-CNN^［32］	WS-DAN^［33］	TASN^［34］	DCL^［35］	TransFG^［36］
Backbone	ResNet-50	VGG-19	VGG-19	ResNet-50	ResNet-50	ResNet-50	ViT_B_16
Top1 err./%	17.45	20.14	20.31	18.13	18.70	18.57	20.43
Top5 err./%	5.20	8.54	8.29	6.18	6.72	7.04	8.97

Table 6. Comparison of the experimental results with fine-grained classification model for fine-grained task on GDIM-CD

Ying LIU, Wei JIANG, Guandian LI, Lei CHEN, Shuang ZHAO. Textile defect recognition network based on label embedding[J]. Optics and Precision Engineering, 2023, 31(10): 1563

Download Citation

Tools

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享