Weak feature defect detection method for LCD screens based on YOLOv5

Table 2. Comparison experiments of introducing Involution layers in different locations of the backbone network
View table
View in Article
Table 2. Comparison experiments of introducing Involution layers in different locations of the backbone network
实验第二层第四层第六层 Precision/% Recall/% FLOPs/G mAP@0.5/%
1 96.4 88.9 15.8 94.5
2 √ 93.5 91.5 5.3 92.1
3 √ 92.5 89.0 6.8 90.5
4 √ 91.8 90.3 16.0 89.6
5 √ √ 87.1 86.3 3.0 90.5
6 √ √ √ 87.0 85.5 3.0 89.9

Table 3. Improved upsampling method
View table
View in Article
Table 3. Improved upsampling method
Model Precision/% Recall/% mAP@0.5/%
YOLO-I 93.5 91.5 92.1
YOLO-IC 94.6 93.0 93.8

Table 4. Performance of model with different attention modules
View table
View in Article
Table 4. Performance of model with different attention modules
注意力机制 Precision/% Recall/% mAP@0.5/%
YOLO-IC 94.6 93.0 93.8
ECA 95.2 92.7 95.8
CA 93.4 94.5 94.2
SA 93.5 93.8 94.7
BiFormer 96.0 95.7 96.5

Table 5. Improved feature fusion
View table
View in Article
Table 5. Improved feature fusion
特征融合方式 Precision/% Recall/% mAP@0.5/%
自顶向下 92.7 83.3 92.1
自底向上 94.8 83.4 93.4
PANet 96.0 95.7 96.5
ASFF 98.4 95.9 97.0
BiFPN 98.6 96.0 97.2

Table 6. Ablation experimental results

View table

View in Article

Table 6. Ablation experimental results

序号	Involution	CARAFE	BiFormer	BiFPN	FLOPs/G	Precision/%	Recall/%	mAP@0.5/%	FPS
A					15.8	96.4	88.9	94.5	99.0
B	√				5.3	93.5	91.5	92.1	131.6
C		√			16.5	94.0	92.7	94.8	116.3
D	√	√			5.9	94.6	93.0	93.8	122.4
E	√		√		6.0	94.2	93.1	92.6	119.0
F		√	√		19.1	94.5	90	93.2	112.5
G	√	√	√		6.1	96.0	95.7	96.5	109.8
H	√	√		√	5.8	94.8	93.5	94.3	113.5
J	√		√	√	5.6	96.2	95.4	94.6	110.3
I	√	√	√	√	6.3	98.6	96.5	97.2	103.1

Table 7. Comparison of grayscale values of missed defects
View table
View in Article
Table 7. Comparison of grayscale values of missed defects
模型漏检数量背景平均灰度值缺陷平均灰度值差值
YOLOv5s 45 184.57 179.75 4.82
YOLO-Mura 6 184.74 182.12 2.62

Table 8. Performance comparison of different algorithms

View table

View in Article

Table 8. Performance comparison of different algorithms

Model	FLOPs/G	Precision/%	Recall/%	mAP@0.5
YOLOv3-Tiny	12.9	84.9	88.2	90.1
YOLOv4s-Mish	20.6	91.6	93.5	95.4
YOLOv5n	4.2	93.6	90.1	93.2
YOLOv5s	15.8	96.4	88.9	94.5
YOLOv7-Tiny	13.0	87.1	86.3	90.5
YOLOv8n	8.0	92.8	92.2	94.8
YOLOv8s	28.4	93.4	92.9	95.3
YOLO-Mura	6.3	98.6	96.5	97.2

Tools

Get Citation

Copy Citation Text

Feng LIN, Yan SHI, Shunlong CHEN, Yinghua LIAO, Lian ZHAO, Li ZHAO, Zemin ZHOU. Weak feature defect detection method for LCD screens based on YOLOv5[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(6): 790

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Jun. 6, 2023

Accepted: --

Published Online: Jul. 30, 2024

The Author Email: Yan SHI (sy71Email@163.com)

DOI:10.37188/CJLCD.2023-0206

Topics

Table 1. LCD Mura defect data set

Table 1. LCD Mura defect data set

Table 2. Comparison experiments of introducing Involution layers in different locations of the backbone network

Table 2. Comparison experiments of introducing Involution layers in different locations of the backbone network

Table 3. Improved upsampling method

Table 3. Improved upsampling method

Table 4. Performance of model with different attention modules

Table 4. Performance of model with different attention modules

Table 5. Improved feature fusion

Table 5. Improved feature fusion

Table 6. Ablation experimental results

Table 6. Ablation experimental results

Table 7. Comparison of grayscale values of missed defects

Table 7. Comparison of grayscale values of missed defects

Table 8. Performance comparison of different algorithms

Table 8. Performance comparison of different algorithms