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

Table 4 Comparative analysis of the Deep Ensemble design with various deep learning architectures with a data split of 60:20:20.

Model	Methods	Validation-loss	Validation-accuracy	Test-accuracy
Our proposed model I (MobileNetV2+InceptionV3)	FT on ADS-ALUF	0.0284	98.96	99.94
Proposed model II (GoogleNet+SqueezeNet)	FT on ADS-ALUF	0.3087	97.72	99.85
AlexNet	Trained on ODS-NPTW	5.4068	8.07	8.46
	PT on ODS-NPTW	0.6312	74.37	74.87
	FT on ODS-PTW	0.3788	95.64	94.14
	FT on ADS-ALUF	0.3230	94.93	94.86
SqueezeNet	Trained on ODS-NPTW	3.1764	5.86	5.53
	PT on ODS-NPTW	0.3806	79.38	79.62
	FT on ODS-PTW	0.2673	86.47	86.33
	FT on ADS-ALUF	0.0646	93.43	92.90
GoogleNet	Trained on ODS-NPTW	3.2394	4.16	4.17
	PT on ODS-NPTW	0.3559	88.68	88.35
	FT on ODS-PTW	0.3937	91.35	90.76
	FT on ADS-ALUF	0.2774	95.71	95.83
MobileNetV2	Trained on ODS-NPTW	3.3572	4.16	4.10
	PT on ODS-NPTW	1.8974	43.98	41.73
	FT on ODS-PTW	0.4196	93.62	94.73
	FT on ADS-ALUF	0.0879	95.95	96.16
InceptionV3	Trained on ODS-NPTW	3.3490	4.03	4.23
	PT on ODS-NPTW	2.3959	58.82	57.94
	FT on ODS-PTW	0.4737	95.51	96.06
	FT on ADS-ALUF	0.2032	97.14	96.68

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