Recognition of rotated images by angle estimation using feature map with CNN
Nishiki Katayama, Satoshi Yamane · 2017
We propose a model that adapts to CNN trained by non-rotated images even for rotated images by evaluating the feature map obtained from the convolution part of CNN. The additional network for rotation angle estimation is able to correct the rotation angle using the feature map in the MNIST data-set. It is possible to cope with the rotation without changing the original network by adding a network to judge.