Application Research for Fusion Model of Pseudolabel and Cross Network

Table 2

Experimental results of cotraining part.

Hyperparameters		Model
		VGG16 [11]	ResNet50 [12] (%)	Transfer network		Multitask network
				VGG16 (%)	ResNet50 (%)	VGG16 (%)	ResNet50 (%)
Revolution	32 × 32	56.04	52.45	53.15	54.55	60.94	60.29
	64 × 64	55.36	53.36	55.44	54.34	61.10	60.80
	72 × 72	56.37	54.71	53.71	54.59	60.25	60.10
	96 × 96	55.22	54.20	55.12	55.42	60.94	60.60
	100 × 100	61.38	59.34	61.29	59.47	60.10	57.43
	128 × 128	59.66	56.65	55.87	54.86	61.45	60.74
Learning rate	0.001	56.74	55.56	55.11	54.27	61.95	61.50
	0.003	61.38	59.34	61.29	59.47	60.10	57.43
	0.01	56.07	55.05	56.37	55.74	55.82	59.74
	0.03	57.42	54.49	56.65	56.57	54.02	50.64
	0.1	46.18	55.32	55.44	55.37	48.36	48.44
	0.3	48.23	46.53	47.61	45.46	47.98	45.55

Better results in terms of recognition rates were highlighted in bold.