A Robust Residual Shrinkage Balanced Network for Image Recognition from Japanese Historical Documents

Table 2

The comparison results with other existing techniques in the Kuzushiji-49 dataset.

Methods	Data descriptions	Balanced accuracy
4-nearest neighbor baseline	Raw data input	86.01%
Keras Simple CNN Benchmark	Raw data input	89.25%
PreActResNet-18	Raw data input	96.64%
	With data augmentation	97.04%
	With data augmentation	97.33%
ResNet-18	Raw data input	96.84%
Proposed-basic	Raw data input	96.89%
Proposed	Raw data input	97.16%