Object detection of steel surface defect based on multi-scale enhanced feature fusion

Shanling LIN; Xueling PENG; Dong WANG; Zhixian LIN; Jianpu LIN; Tailiang GUO

doi:10.37188/OPE.20243207.1075

Optics and Precision Engineering, Volume. 32, Issue 7, 1075(2024)

Object detection of steel surface defect based on multi-scale enhanced feature fusion

Shanling LIN^1,2, Xueling PENG^1,2, Dong WANG^1,2, Zhixian LIN^1,2,3, Jianpu LIN^1,2、*, and Tailiang GUO^2,3

Author Affiliations

¹School of Advanced Manufacturing， Fuzhou University， Quanzhou362252， China

²China Fujian Photoelectric Information Science and Technology Innovation Laboratory， Fuzhou350116， China

³School of Physics and Information Engineering， Fuzhou University， Fuzhou50116， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Network structure of EFSSD-MV2 algorithm

Download full size

View in Article

Fig. 2. Adaptive Weighted Fusion（AWF） module

Download full size

View in Article

Fig. 3. Spatial Feature Enhancement（SFE） module

Download full size

View in Article

Fig. 4. Comparison of algorithm detection accuracy before and after improvement

Download full size

View in Article

Fig. 5. Comparison of class activation mapping output before and after algorithm improvement

Download full size

View in Article

Fig. 6. Comparison of sample detection results before and after algorithm improvement

Download full size

View in Article

Table 1. Ablation experiments for EFSSD-MV2 algorithm（1 M=10⁶）

View table

View in Article

Table 1. Ablation experiments for EFSSD-MV2 algorithm（1 M=10⁶）

方法	mAP/%	mAPITSS/%	参数量/M	计算量（MFLOPs）
SSD-MV2	73.66	77.41	1.92	568.68
FSSD-MV2	76.11	78.78	2.24	808.16
FSSD-MV2AWF	77.33	79.61	2.26	808.66
FSSD-MV2SFE	77.18	79.98	2.34	952.54
FSSD-MV2EFF	77.85	80.47	2.36	952.67

Table 2. Comparison of experimental results with other lightweight target detection algorithms

View table

View in Article

Table 2. Comparison of experimental results with other lightweight target detection algorithms

方法	AP/%						mAP /%	模型大小 /MB	fps/ （frame·s^-1）
方法	裂纹	补丁	杂质	点蚀	结疤	刮痕	mAP /%	模型大小 /MB	fps/ （frame·s^-1）
YOLOv3-tiny^［18］	36.09	88.91	73.11	65.41	59.77	83.76	67.84	33.14	60.7
YOLOv4-tiny^［19］	36.46	87.67	67.37	79.48	58.26	81.67	68.49	22.48	64.1
YOLOv5n^［20］	38.40	90.70	80.90	85.50	65.60	93.20	75.72	14.25	91.8
YOLOv6n^［21］	39.30	92.60	79.50	85.20	60.90	95.80	75.55	9.29	86.8
YOLOv7-tiny^［22］	30.00	91.60	83.80	85.30	65.60	92.90	74.87	11.75	85.0
YOLOX-nano^［23］	43.32	92.14	79.65	83.10	66.28	96.18	76.78	7.23	54.8
SSDLite^［24］	42.51	92.93	74.85	84.84	68.22	80.85	74.03	24.98	68.0
NanoDet^［25］	3.00	87.40	54.40	68.60	30.00	58.30	50.28	10.98	74.8
FastestDet^［26］	39.70	86.80	64.10	79.40	75.40	83.10	71.42	1.08	97.7
MutualGuide^［27］	51.03	86.72	81.73	84.78	66.61	89.47	76.72	10.96	68.6
EFSSD-MV2	55.13	91.94	79.54	88.24	74.57	93.42	80.47	9.27	66.4

Table 3. Effect of hyperparameter k on mAP metrics of EFSSD-MV2 algorithm
View table
View in Article
Table 3. Effect of hyperparameter k on mAP metrics of EFSSD-MV2 algorithm
k mAP_ATSS mAP_ITSS
5 76.55 78.44
10 79.10 79.38
15 79.53 79.44
20 79.04 79.73
25 79.04 80.03
30 79.81 80.47
35 79.70 79.98
40 79.06 79.33
45 78.72 78.26
50 78.90 78.18

Tools

Get Citation

Copy Citation Text

Shanling LIN, Xueling PENG, Dong WANG, Zhixian LIN, Jianpu LIN, Tailiang GUO. Object detection of steel surface defect based on multi-scale enhanced feature fusion[J]. Optics and Precision Engineering, 2024, 32(7): 1075

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites