Defect Detection for Solar Cells using Dense Backbone Network Algorithm

Fig. 2. Defect image classification. (a) Single crystal silicon crack; (b) polysilicon crack; (c) single crystal silicon finger-interruption; (d) polysilicon finger-interruption

Download full size

Fig. 3. YOLOv4 network structure

Download full size

Fig. 4. Detection results of the proposed algorithm

Download full size

Fig. 5. Detection results of the YOLOv4 algorithm

Download full size

Fig. 6. AP value comparison

Download full size

Fig. 7. Loss value curve

Download full size

Fig. 8. F1 value of crack

Download full size

Fig. 9. F1 value of finger-interruption

Download full size

Table 1. DenseNet121 model parameters

View table

Table 1. DenseNet121 model parameters

Layer	Module	Stride	Output size
Convolution	7×7 Conv	2	800×800
Pooling	7×7 max pooling	2	400×400
Dense block	（1×1 Conv）×6 （3×3 Conv）×6		400×400
Transition layer	1×1 Conv	1	400×400
Transition layer	2×2 average pooling	2	200×200
Dense block	（1×1 Conv）×12 （3×3 Conv）×12		200×200
Transition layer	1×1 Conv	1	200×200
Transition layer	2×2 average pooling	2	100×100
Dense block	（1×1 Conv）×24 （3×3 Conv）×24		100×100
Transition layer	1×1 Conv	1	100×100
Transition layer	2×2 average pooling	2	50×50
Dense block	（1×1 Conv）×16 （3×3 Conv）×16		50×50
Classification layer			1×1

Table 2. Model parameter setting
View table
Table 2. Model parameter setting
Parameter Value
Input size 800×800
Freeze training epoch 100
Freeze training learning rate 0.001
Unfreeze training epoch 100
Unfreeze training learning rate 0.0001
Label_smoothing 0.005
N_T∶N_V 4∶1

Table 3. Comparison of detection results of different algorithms

View table

Table 3. Comparison of detection results of different algorithms

Algorithm	Defect	TP	FP	FN
Proposed algorithm	Finger-interruption	3148	551	63
Proposed algorithm	crack	1865	325	51
YOLOv4	Finger-interruption	3027	711	184
YOLOv4	crack	1772	474	144
Efficientnet-YOLOv3	Finger-interruption	2907	912	304
Efficientnet-YOLOv3	crack	1645	864	271
YOLOv4-tiny	Finger-interruption	2953	735	258
YOLOv4-tiny	crack	1659	589	257
Faster-rcnn	Finger-interruption	2775	17236	441
Faster-rcnn	crack	1721	10879	195
YOLOv5	Finger-interruption	3097	596	114
YOLOv5	crack	1827	412	89

Table 4. Ablation experiment

View table

Table 4. Ablation experiment

Ghostnet	Mobilenetv1	CSPDarkNet53	DenseNet121	NMS	Softer-NMS	mAP /%	Speed /（frame·s^-1）	TP	FP	FN
		√		√		87.14	22.73	4799	1185	328
			√	√		90.37	30.16	4908	1002	219
		√			√	90.24	19.65	4926	1096	206
			√		√	93.08	27.35	5013	876	114
	√				√	81.95	28.42	4611	1328	516
√					√	84.26	28.87	4685	1127	442

Table 5. Comparison of model indicators

View table

Table 5. Comparison of model indicators

Algorithm	mAP /%	Speed/（frame·s^-1）	Param /MB
Proposed algorithm	93.08	27.35	160.4
YOLOv4	87.14	22.73	244.6
Efficientnet-YOLOv3	84.02	23.20	154.1
YOLOv4-tiny	83.36	113.5	22.6
Faster-rcnn	63.48	12.06	108.7
YOLOv5	90.21	19.47	335.3

Tools

Get Citation

Copy Citation Text

Zheng Tang, Huilin Zhang, Lixin Ma. Defect Detection for Solar Cells using Dense Backbone Network Algorithm[J]. Laser & Optoelectronics Progress, 2023, 60(14): 1412003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites