A Novel Data Analytics Oriented Approach for Image Representation Learning in Manufacturing Systems

Table 2

The comparison with state-of-the-art self-supervised methods in ImageNet classification, evaluated by linear probing. “Params” and “Architecture” column shows the basic features of the corresponding method. “Top1 Acc.” and “Top5 Acc.” are reported by a linear classification on the ImageNet dataset, after models are pretrained with self-supervision.

Method	Params	Architecture	Top1 Acc. (%)	Top5 Acc. (%)
CMC	94 M	ResNet50 ×2	70.6	89.7
CPC v2	305 M	ResNet161	71.5	90.1
BYOL	375 M	ResNet50 ×4	78.6	94.2
SimCLR	375 M	ResNet50 ×4	76.5	93.2
MoCo v3	304 M	ViT-L/16	77.6	—
MoCo v3	304 M	ViT-BN-L/7	81.0	—
TriLFrame (ours)	75 M	ResNet18 + transformer	75.6	92.1
TriLFrame (ours)	265 M	ResNet50 + transformer	78.3	93.6
TriLFrame (ours)	485 M	ResNet101 + transformer	81.2	94.7