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

Table 1

The result of ablation studies on TriLFrame architecture. The model is pretrained with contrastive learning and then fine-tuned with supervised learning classification task on the ILSVRC ImageNet dataset. “Random Init.” that represents random initialization is used for setting up. “Remove transEnc()” that represents the transformer of TriLFrame is removed; patch embeddings are aggregated without transformer encoder.

Encoder	Self-Sup. (ILSVRC)		Sup. (ILSVRC)
	Setting	Top1 Acc.	Top1 Acc.
ResNet18	Random Init.	—	54.1
ResNet18	Remove	68.8	66.8
ResNet18	TriLFrame	80.9	75.6
ResNet50	TriLFrame	83.7	78.3
ResNet101	TriLFrame	88.5	81.2