Enhancing the Power of CNN Using Data Augmentation Techniques for Odia Handwritten Character Recognition

Table 10

Performance comparison on handwritten ISI image and Odia numeral databases.

Database	Work reference		Features	Recognition classifier	Recognition accuracy (%)
ISI image database of handwritten oriya numerals	[11]		Binary external symmetry axis constellation	Random forest	98.44
	[41]		Scalar and stroke information	HMM	90.50
	[42]		LU decomposition of matrix factors	Backpropagation ANN	85.30
	[14]		Gradient, curvature	Low complexity neural network	98 (gradient); 94 (curvature)
	[16]		Slantlet transform, stockwell transform and gabor	k-NN	95.04 (slantlet); 98.08 (stockwell)
	Proposed work (without augmentation)		Features obtained from convolutional layer	CNN	97.39
	Proposed work (with augmentation)	Translation	Features obtained from convolutional layer	CNN	98.60
		Rotation			98.00
		Scaling			97.79
		Elastic deformation			97.69
		Noise			97.60
		Color inversion			98.00