Deep ensemble approach for pathogen classification in large-scale images using patch-based training and hyper-parameter optimization

Table 3 Comparative analysis of the Deep Ensemble design with various Deep learning architectures with a data split of 70:20:10.

Model	Methods	Validation-loss	Validation-accuracy	Test-accuracy
Proposed model I (MobileNetV2+InceptionV3)	FT on ADS-ALUF	0.0275	99.41	99.91
Proposed model II (GoogleNet+SqueezeNet)	FT on ADS-ALUF	0.0953	97.14	99.79
AlexNet	Trained on ODS-NPTW	4.1254	13.53	11.98
	PT on ODS-NPTW	0.8999	75.47	76.43
	FT on ODS-PTW	0.2375	93.10	93.10
	FT on ADS-ALUF	0.1056	96.23	96.09
SqueezeNet	Trained on ODS-NPTW	3.1779	6.51	8.07
	PT on ODS-NPTW	0.3370	79.83	82.16
	FT on ODS-PTW	0.3102	87.51	87.89
	FT on ADS-ALUF	0.2614	94.60	93.62
GoogleNet	Trained on ODS-NPTW	4.1299	2.93	1.95
	PT on ODS-NPTW	0.3924	87.77	88.41
	FT on ODS-PTW	0.5174	90.83	90.63
	FT on ADS-ALUF	0.2897	95.64	93.75
MobileNetV2	Trained on ODS-NPTW	3.1169	5.27	5.86
	PT on ODS-NPTW	2.1408	44.05	39.84
	FT on ODS-PTW	0.2569	93.10	92.45
	FT on ADS-ALUF	0.0212	97.85	97.79
InceptionV3	Trained on ODS-NPTW	3.4025	1.69	0.91
	PT on ODS-NPTW	2.4302	56.93	57.03
	FT on ODS-PTW	0.4128	96.62	96.74
	FT on ADS-ALUF	0.0446	98.39	97.92

Here, ADS, ODS, PTW, NPTW, ALUF, PT, FT stands for augmented dataset, original dataset, pre-trained weights, no pre-trained weights, all layers un-frozen, parameter-tuning, fine-tuned

ISSN: 1471-2105