Lightweight Tilted Droplet Identification Method Based on Improved YOLOv8

Zhaojin Wu; Jun Wang; Zhaoliang Cao

doi:10.3788/LOP242472

Laser & Optoelectronics Progress, Volume. 62, Issue 12, 1237009(2025)

Lightweight Tilted Droplet Identification Method Based on Improved YOLOv8

Zhaojin Wu^1,3, Jun Wang^2、*, and Zhaoliang Cao^1,3

¹Key Laboratory of Intelligent Optoelectronic Devices and Chips of Jiangsu Higher Education Institutions, School of Physical Science and Technology, Suzhou University of Science and Technology, Suzhou 215009, Jiangsu , China

²School of Electronic & Information Engineering, Suzhou University of Science and Technology, Suzhou 215009, Jiangsu , China

³Advanced Technology Research Institute of Taihu Photon Center, School of Physical Science and Technology, Suzhou University of Science and Technology, Suzhou 215009, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(18)

Fig. 1. Light-YOLOv8OBB network architecture

Download full size

Fig. 2. General flow chart of tilted droplet detection and localization

Download full size

Fig. 3. Illustration of the Ghost Bottleneck^[17]. (a) Stride is 1; (b) stride is 2

Download full size

Fig. 4. C2f module and C2f-light module. (a) Illustration of C2f module; (b) illustration of C2f-light module

Download full size

Fig. 5. Illustration of GhostConv module

Download full size

Fig. 6. Illustration of three different module structures. (a) GSConv module; (b) GSBottleneck module; (c) VoV-GSCSP module

Download full size

Fig. 7. Overview of CBAM module

Download full size

Fig. 8. Flow chart of post-processing model for tilted droplet image

Download full size

Fig. 9. Data samples

Download full size

Fig. 10. Comparison of mAP@0.5：0.95 values between original and improved algorithm

Download full size

Fig. 11. Visualization of model detection and localization results

Download full size

Table 1. Parameters for training
View table
Table 1. Parameters for training
Parameter Value Parameter Value
Pretrained False Mosaic 0.8
Epoch 300 Mixup 0.7
Imgsize 640×640 HSV_S 0.8
Batch 4 HSV_V 0.8
IoU 0.7 Optimizer Adam

Table 2. Experimental results of replacing convolutional module in backbone
View table
Table 2. Experimental results of replacing convolutional module in backbone
Conv module Params /10⁶ FLOPs /10⁹ mAP@0.5：0.95 /%
C2f 3.1 8.3 69.2
C3 2.6 7.1 70.7
C3Ghost 2.6 6.8 73.9
C2f-light 2.2 6.1 75.8

Table 3. Experimental results of lightweight neck network
View table
Table 3. Experimental results of lightweight neck network
Neck Params /10⁶ FLOPs /10⁹ mAP@0.5：0.95 /%
Base neck 3.1 8.3 69.2
Ghost neck 2.3 6.6 73.2
Slim-Neck 2.2 6.4 72.7

Table 4. Comparison results of different attention mechanisms
View table
Table 4. Comparison results of different attention mechanisms
Module Params /10⁶ FLOPs /10⁹ mAP@0.5：0.95 /%
Base 3.1 8.3 69.2
Base+CA 2.6 7.1 71.3
Base+SE 2.6 7.2 72.2
Base+ECA 2.6 7.3 71.5
Base+CBAM 2.7 7.2 73.8

Table 5. Comparison results of the ablation experiments
View table
Table 5. Comparison results of the ablation experiments
Module Params /10⁶ FLOPs /10⁹ mAP@0.5：0.95 /%
Base 3.1 8.3 69.2
Base+C2f-light 2.2 6.1 75.8
Base+C2f-light+Slim-Neck 1.8 5.2 73.9
Base+C2f-light+Slim-Neck+CBAM 1.9 5.4 76.7

Table 6. Comparison of experimental results before and after improvement
View table
Table 6. Comparison of experimental results before and after improvement
Algorithm Params /10⁶ FLOPs /10⁹ mAP@0.5：0.95（1） /% mAP@0.5：0.95（2）/% Inference time /ms
YOLOv8n 3.2 8.7 53.6 35.7 21.3
YOLOv8n-OBB 3.1 8.3 69.2 47.8 20.5
Light-YOLOv8OBB 1.9 5.4 76.7 67.3 16.1

Table 7. Comparison results of different algorithms on tilted droplet dataset

View table

Table 7. Comparison results of different algorithms on tilted droplet dataset

Algorithm	Params /10⁶	FLOPs /10⁹	mAP@0.5：0.95 /%	Inference time /ms
YOLOv5n	2.7	7.7	49.8	18.9
YOLOv9t-OBB	2.0	8.2	70.8	26.7
Gliding vertex	41.2	121.5	65.8	43.5
RoI transformer	122.7	55.1	79.6	50.7
YOLOv11n-OBB	2.7	6.9	73.2	22.3
Oriented R-CNN	121.6	41.1	72.8	39.8
Proposed	1. 9	5.4	76.7	16.1

Tools

Get Citation

Copy Citation Text

Zhaojin Wu, Jun Wang, Zhaoliang Cao. Lightweight Tilted Droplet Identification Method Based on Improved YOLOv8[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1237009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites