The Performance Research of the Data Augmentation Method for Image Classification

Table 3

Comparison of the performance of the five augmentation methods and natural input images in the three datasets. The training and testing accuracy of color augmentation is average by the color and contrast transformation method.Again, the traditional enhancement value comes from the average of rotations, flips, and shifts.

	Network	Color augmentation		Origin		Traditional augmentation
		Train	Test	Train	Test	Train	Test
ImageNet	AlexNet	0.918	0.662	0.933	0.663	0.664	0.647
	VGG16	0.952	0.591	0.946	0.601	0.672	0.626
	InceptionV3	0.853	0.589	0.893	0.586	0.598	0.593
VOC	AlexNet	0.865	0.553	0.945	0.554	0.676	0.577
	VGG16	0.962	0.497	0.923	0.483	0.628	0.558
	InceptionV3	0.966	0.454	0.962	0.470	0.696	0.510
Pests	AlexNet	0.994	0.809	0.996	0.805	0.972	0.846
	VGG16	0.991	0.751	0.992	0.758	0.955	0.825
	InceptionV3	0.995	0.779	0.980	0.776	0.913	0.812