Bulk Damage Point Detection in Crystals Based on Improved YOLOv8

Haojie Feng; Jinfang Shi; Rong Qiu; Qiang Zhou; Jianxin Wang; Decheng Guo; Qing Wang

doi:10.3788/LOP240590

Laser & Optoelectronics Progress, Volume. 61, Issue 22, 2212004(2024)

Bulk Damage Point Detection in Crystals Based on Improved YOLOv8

Haojie Feng¹, Jinfang Shi^1、*, Rong Qiu², Qiang Zhou², Jianxin Wang², Decheng Guo², and Qing Wang²

¹School of Manufacturing Science and Engineering, Southwest University of Science and Technology, Mianyang 621010, Sichuan , China

²Joint Laboratory for Extreme Conditions Matter Properties, School of Mathematics and Physics, Southwest University of Science and Technology, Mianyang 621010, Sichuan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Enhanced network structure of the YOLOv8 Model

Download full size

Fig. 2. CBAM attention mechanism structure and its details. (a) CBAM structure; (b) channel attention mechanism; (c) spatial attention mechanism

Download full size

Fig. 3. Optical path diagram of DKDP crystal laser damage testing experimental system

Download full size

Fig. 4. Some typical images of crystal bulk damage

Download full size

Fig. 5. Morphology of damage points under microscope. (a) Strong demage point; (b) weak demage piont

Download full size

Fig. 6. Prediction results comparison. (a) (b) The comparison of small-scale damage detection results; (c) (d) the comparison of large-scale damage detection results

Download full size

Fig. 7. Heatmap of predicted results

Download full size

Table 1. Dataset details

View table

Table 1. Dataset details

Classes	Number	Area /pixel²	Cause	Feature description
WD	2334	4‒10	High-power laser irradiation	Tiny dim light spots
SD	2207	10‒80	High-power laser multiple irradiation	Small-scale light spot with aperture at the edge
LSFD	1405	30‒300	Scattered light from laser on the surface of optical components	Multi-scale strip light spot
BPFD	617	20‒40	Scattered laser light at the damage point destroys the camera sensor	Small-scale light spots with irregular shapes

Table 2. Ablation experiment

View table

Table 2. Ablation experiment

A	B	C	mAP@50% /%	mAP@95% /%	Param /10⁶	GFLOPs /10⁹	Model size /MB
			66.9	44.5	3.011	8.197	6.3
√			69.7	45.2	3.165	8.330	6.6
	√		68.5	45.7	2.938	12.511	6.4
		√	68.9	47.2	3.011	8.197	6.3
√	√		67.9	44.5	5.300	58.400	11.2
√		√	69.7	46.0	3.165	8.330	6.6
	√	√	67.4	45.0	2.938	12.511	6.4
√	√	√	70.0	47.6	5.307	58.686	40.2

Table 3. Comparison experiment

View table

Table 3. Comparison experiment

Model	mAP@50% /%	mAP@95% /%	Param /10⁶	GFLOPs /10⁹	Mode size /MB
RetinaNet	64.7	42.5	37.96	203.0	306
SSD	64.4	42.2	24.14	215.0	194.5
Faster RCNN	66.7	41.5	41.75	178.0	336
YOLOv3-tiny	56.1	39.3	12.13	19.04	24.4
YOLOv5s	65.6	46.6	9.12	24.05	18.5
YOLOX-s	65.3	44.3	8.94	31.78	72.1
YOLOv8-OCD	70.0	47.6	5.30	58.68	40.2

Table 4. Model inference speed
View table
Table 4. Model inference speed
Model Inference time /ms FPS
YOLOv8 22.5 44.44
Improved YOLOv8 105 9.52

Tools

Get Citation

Copy Citation Text

Haojie Feng, Jinfang Shi, Rong Qiu, Qiang Zhou, Jianxin Wang, Decheng Guo, Qing Wang. Bulk Damage Point Detection in Crystals Based on Improved YOLOv8[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2212004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites