Industrial Image Anomaly Detection Using Bias-Reduced Coupled Hyperspheres

Method		RIAD^［3］	ADTR^［6］	CutPaste^［7］	DRAEM^［8］	CS-Flow^［9］	PatchCore^［10］	CFA^［11］	Proposed
Texture	Carpet	84.2/96.3	100.0/98.7	93.9/98.3	97.0/95.5	100.0/-	98.7/99.0	97.3/-	100.0/99.0
	Grid	99.6/98.8	97.5/95.0	100/97.5	99.9/99.7	99.0/-	98.2/98.7	99.2/-	98.7/98.6
	Leather	100.0/99.4	100.0/98.1	100/99.5	100.0/98.6	100.0/-	100.0/99.3	100.0/-	100.0/99.1
	Tile	98.7/89.1	100.0/93.8	94.6/90.5	99.6/99.2	100.0/-	98.7/95.6	99.4/-	99.8/98.0
	Wood	93.0/85.8	99.9/91.2	99.1/95.5	99.1/96.4	100.0/-	99.2/95.0	99.7/-	100.0/94.9
	Average	95.1/93.9	99.5/94.7	97.5/96.3	99.1/97.9	99.8/-	99.0/97.5	99.1/97.2	99.7/97.9
Object	Bottle	99.9/98.4	100.0/98	98.2/97.6	99.2/99.1	99.8/-	100.0/98.6	100.0/-	100.0/98.5
	Cable	81.9/84.2	92.5/96.8	81.2/90.0	91.8/94.7	99.1/-	99.5/98.4	99.8/-	100.0/98.8
	Capsule	88.4/92.8	93.1/99.1	98.2/97.4	98.5/94.3	97.1/-	98.1/98.8	97.3/-	97.7/98.8
	Hazelnut	83.3/96.1	100.0/98.6	98.3/97.3	100.0/99.7	99.6/-	100.0/98.7	100.0/-	100.0/98.6
	Metal nut	88.5/92.5	94.9/97	99.9/93.1	98.7/99.5	99.1/-	100.0/98.4	100.0/-	100.0/98.4
	Pill	83.8/95.7	92.1/98.3	94.9/95.7	98.9/97.6	98.6/-	96.6/97.4	97.9/-	98.6/98.8
	Screw	84.5/98.8	94/99.3	88.7/96.7	93.9/97.6	97.6/-	98.1/99.4	97.3/-	98.2/99.4
	Toothbrush	100.0/98.9	93.1/98.5	99.4/98.1	100.0/98.1	91.9/-	100.0/98.7	100.0/-	100.0/98.3
	Transistor	90.9/87.7	97.6/97.9	96.1/93.0	93.1/90.9	99.3/-	100.0/96.3	100.0/-	100.0/98.4
	Zipper	98.1/97.8	95.8/97.2	99.9/99.3	100.0/98.8	99.7/-	99.4/98.8	99.6/-	100.0/98.8
	Average	89.9/94.3	95.3/98.1	95.5/95.8	97.4/97.0	98.2/-	99.2/98.4	99.2/98.6	99.5/98.7
Total average		91.7/94.2	96.4/97.2	96.2/96.0	98.0/97.3	98.7/-	99.1/98.1	99.2/98.2	99.5/98.4

Table 2. Ablation experimental results of model fine-tuning
View table
Table 2. Ablation experimental results of model fine-tuning
Backbone Model fine-tuning Value /%
CFA Proposed
ResNet18 ‒ 98.7/97.6 99.0/97.7
√ 99.1/97.8 99.4/98.0
WideResNet50 ‒ 99.2/98.2 99.4/98.3
√ 99.4/98.3 99.5/98.4

Table 3. Ablation experimental results of anomaly generation strategy

View table

Table 3. Ablation experimental results of anomaly generation strategy

Generation strategy	Anomaly shape			Anomaly fill		AUROC /%
Generation strategy	Rectangle	Bézier shape	Perlin noise	CutPaste^［7］	External dataset	AUROC /%
‒		‒	‒	‒	‒	99.0/97.7
CutPaste^［7］	√	‒	‒	√	‒	98.7/97.5
DRAEM^［8］	‒	‒	√	‒	√	98.8/97.6
①	√	‒	‒	√	√	99.1/97.7
②	‒	√	‒	√	‒	99.3/97.9
Proposed	‒	√	‒	√	√	99.4/98.0

Table 4. Ablation study results of other improvement methods
View table
Table 4. Ablation study results of other improvement methods
Baseline Baseline+① Baseline+② Baseline+③
Texture 99.1/97.2 99.3/97.5 99.0/97.0 99.4/97.4
Object 99.2/98.6 99.3/98.6 99.4/98.7 99.3/98.7
Total 99.2/98.2 99.3/98.3 99.3/98.2 99.4/98.3

Tools

Get Citation

Copy Citation Text

Shuai Zhang, Meiju Liu. Industrial Image Anomaly Detection Using Bias-Reduced Coupled Hyperspheres[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1215006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites