A New Transfer Learning Ensemble Model with New Training Methods for Gear Wear Particle Recognition

Table 3

Classification accuracy when fine-tuned CNN is used for feature extraction for further input to support vector machine (SVM) classification.

		DCNN + SVM	VGG19 + SVM	GoogLeNet + SVM	(VGG19 + GoogLeNet) + SVM	LCNNE + SVM
1	Accuracy	94.63%	97.35%	99.56%	96.30%	99.26%
2	Accuracy	96.57%	98.95%	99.68%	96.80%	99.76%
3	Accuracy	97.18%	99.13%	97.45%	96.77%	99.42%
4	Accuracy	93.35%	97.62%	97.99%	95.61%	99.74%
5	Accuracy	93.69%	98.79%	97.70%	94.66%	99.20%
6	Accuracy	96.32%	99.70%	99.39%	97.63%	99.82%
7	Accuracy	95.51%	99.81%	99.94%	96.97%	99.96%
8	Accuracy	95.39%	99.16%	99.34%	96.41%	99.66%
9	Accuracy	98.14%	96.34%	97.66%	96.78%	99.54%
10	Accuracy	96.76%	99.49%	99.14%	95.83%	99.90%
Mean accuracy		95.75%	98.63%	99%	96.38%	99.63%
Standard deviation		1.54%	1.15%	0.97%	0.83%	0.26%