Autofocus Method Based on Deep Learning in the Visual Measurement System

Bowen Zheng; Shaojin Liu; Chengwu Shen; Jianrong Li; Yan Han; Haoyang Sun

doi:10.3788/LOP240938

Laser & Optoelectronics Progress, Volume. 61, Issue 24, 2411003(2024)

Autofocus Method Based on Deep Learning in the Visual Measurement System

Bowen Zheng^1,2,3, Shaojin Liu^1,2,3、*, Chengwu Shen^1,2,3, Jianrong Li^1,2,3, Yan Han^1,2,3, and Haoyang Sun^1,2,3

Author Affiliations

¹Changchun Institute of Optics Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, Jilin , China

²Relative Pose Measurement Laboratory, Changchun Institute of Optics Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, Jilin , China

³University of Chinese Academy of Sciences, Beijing 100049, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Visual measurement system composition

Download full size

Fig. 2. Model training and prediction process

Download full size

Fig. 3. Partial dataset images and their real defocus distance

Download full size

Fig. 4. Decline of training set loss and validation set loss

Download full size

Fig. 5. Defocusing distance of model prediction in test set

Download full size

Fig. 6. Single defocus image sequence sharpness evaluation curve

Download full size

Fig. 7. Schematic diagram of focus sub-region division

Download full size

Fig. 8. Two-step focusing algorithm flow chart

Download full size

Fig. 9. Multi-depth-of-field target scene used in the experiment

Download full size

Fig. 10. Focus error statistics diagram. (a) Group A; (b) group B; (c) group C

Download full size

Table 1. Main imaging parameters of vision measurement system
View table
Table 1. Main imaging parameters of vision measurement system
Parameter Value
Focal distance/m 0.15‒10.00
Resolution 2592（H）×2048（V）
Pixel dimension/μm 4.8

Table 2. Network structure used to predict defocus distance

View table

Table 2. Network structure used to predict defocus distance

Type	Convolutional kernel size or stride	Output size
Input	‒	224×224×3
Conv2D	3×3/2	112×112×24
Max pool	3×3/2	56×56×24
Stage 1	1×unit for spatial down sampling^［15］/2	28×28×116
Stage 1	3×basic unit^［15］/1	28×28×116
Stage 2	1×unit for spatial down sampling/2	14×14×232
Stage 2	7×basic unit/1	14×14×232
Stage 3	1×unit for spatial down sampling/2	7×7×464
Stage 3	3×basic unit/1	7×7×464
Conv2D	1×1/1	7×7×1024
Global pool	7×7/1	1×1×1024
FC 1	‒	1×1×1024
FC 2	‒	1×1×256
FC 3	‒	1×1×1

Table 3. Deep learning training and prediction parameters

View table

Table 3. Deep learning training and prediction parameters

Parameter	Value
Training set images	43354
Validation set images	12386
Test set images	6193
Computer disposition	NVDIA GeForce RTX 3070 Ti
Training environment	PyTorch
Batch size	256
Epochs	200
Training duration /min	312
Predicting duration /ms	23
Model memory /M	6.04

Table 4. Single frame target image defocus distance prediction method
View table
Table 4. Single frame target image defocus distance prediction method
Focus sub-region Region defocus distance Single frame target image defocus distance D
w₁ 2.12 2.10
w₂ 2.69
w₃ 1.62
w₄ 2.09

Table 5. Proposed method is compared with traditional autofocus method

View table

Table 5. Proposed method is compared with traditional autofocus method

Group	Focus start position /V	Proposed method			Traditional autofocus method			Best focus positionV_f /V
Group	Focus start position /V	Detect focus position V_D /V	Focus time /s	Lens adjustment times	Detect focus position V_D /V	Focus time /s	Lens adjustment times	Best focus positionV_f /V
A	35	44.7	0.602	2	44.8	2.312	15	45.2
	38	45.1	0.583	2	44.6	2.905	18
	40	45.0	0.591	2	44.8	2.133	15
B	35	41.5	0.305	2	40.8	1.639	16	42.3
	38	41.4	0.289	2	41.6	1.234	12
	40	41.6	0.337	2	40.4	1.459	14
C	35	41.3	0.293	2	41.2	1.236	12	41.5
	38	41.1	0.294	2	41.2	1.744	16
	40	41.1	0.352	2	41.2	1.115	11

Table 6. Focusing error and focusing stability of the proposed method and the traditional autofocus method
View table
Table 6. Focusing error and focusing stability of the proposed method and the traditional autofocus method
Group Average focusing error /V Focusing stability
Proposed method Traditional autofocus method Proposed method Traditional autofocus method
A 0.281 0.381 0.192 0.218
B 0.707 1.646 0.133 0.507
C 0.381 0.262 0.117 0.192

Tools

Get Citation

Copy Citation Text

Bowen Zheng, Shaojin Liu, Chengwu Shen, Jianrong Li, Yan Han, Haoyang Sun. Autofocus Method Based on Deep Learning in the Visual Measurement System[J]. Laser & Optoelectronics Progress, 2024, 61(24): 2411003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites