Surface-Defect Detection Algorithm for Aluminum Profiles Based on CDA-YOLOv8

Table 1. Experimental parameter configuration
View table
Table 1. Experimental parameter configuration
Parameter Value
Image size 640×640
Optimizer SGD
Batch_size 16
Epoch 300
Momentum 0.937
Learning rate 0.01
weight decay 0.0005

Table 2. Results of a comparative experiment on downsampling
View table
Table 2. Results of a comparative experiment on downsampling
Model Params /10⁶ FLOPs /10⁹ mAP@0.5 /%
Conv （Baseline） 11.1 28.5 83.7
CG Block 10.6 27.1 84.8
SPDConv 15.8 39.8 83.7
ADown 10.0 25.7 84.5
LDConv 10.6 27.7 81.6
WaveletPool 9.9 25.7 84.1

Table 3. Comparison experiment results of C2f module
View table
Table 3. Comparison experiment results of C2f module
Model Params /10⁶ FLOPs /10⁹ mAP@0.5 /%
C2f （baseline） 11.1 28.5 83.7
C2f_DWR 10.6 27.1 85.1
C2f_DySnakeConv 14.4 33.8 83.9
C2f_RFCAConv 11.2 29.3 84.9
C2f_MLCA 11.1 28.5 84.3
C2f_PPA 14.3 34.9 84.2

Table 4. Comparison experiment results of small object detection layer
View table
Table 4. Comparison experiment results of small object detection layer
Model P2 ASFP2 Params /10⁶ FLOPs /10⁹ mAP@0.5 /%
CG Block+C2f_DWR ‒ ‒ 11.3 29.0 86.0
CG Block+C2f_DWR √ ‒ 10.8 36.9 86.0
CG Block+C2f_DWR ‒ √ 9.2 34.6 88.1

Table 5. Ablation experiments

View table

Table 5. Ablation experiments

Model	CGBlock	C2f_DWR	ASFP2	Params /10⁶	FLOPs /10⁹	P /%	R /%	mAP@0.5 /%
1	–	–	–	11.1	28.5	88.3	78.0	83.7
2	√	–	–	11.8	30.3	87.3	80.1	84.8
3	–	√	–	10.6	27.1	88.8	80.6	85.1
4	–	–	√	8.9	34.1	86.8	80.7	85.3
5	√	√	–	11.3	29.0	87.7	83.0	86.0
6	√	–	√	9.6	35.9	86.3	83.0	86.6
7	–	√	√	8.5	32.8	91.0	78.9	85.6
8	√	√	√	9.2	34.6	93.4	80.4	88.1

Table 6. Comparison of detection results of each defect category

View table

Table 6. Comparison of detection results of each defect category

Class	AP /%
Class	YOLOv8s	CDA-YOLOv8
Non-conductive	93.8	96.9
Stains	41.5	47.2
Corner leakage	98.4	98.5
Spray flow	88.8	92.4
Paint bubbles	56.2	71.7
crutches	69.1	81.5
Mixed colors	99.5	99.5
Orange peel	97.7	98.6
Leakage	98.0	98.4
Pitting	93.5	96.7

Table 7. Comparison experimental results
View table
Table 7. Comparison experimental results
Model P /% R /% mAP@0.5 /%
Faster R-CNN 77.8 72.5 75.8
YOLOv5s 85.5 79.8 82.3
YOLOv8s 88.3 78.0 83.7
YOLO9s 87.8 78.2 84.9
YOLOv10s 83.9 76.8 82.8
CDA-YOLOv8 93.4 80.4 88.1

Tools

Get Citation

Copy Citation Text

Yawei Zhao, Geng Sun, Hongjie Wang, Haonan Hu. Surface-Defect Detection Algorithm for Aluminum Profiles Based on CDA-YOLOv8[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1637007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites