Hello, why does the text recognition method based on deep learning still need graying?

[htaoruan]

Hello, why does the text recognition method based on deep learning still need graying? In mmocr, if I want to make the image input to the network non-gray image, how should I modify it?

[gaotongxiao]

1. In MMOCR, only CRNN takes grayscale images. Other architectures such as ABINet works on RGB images.
2. You will need to modify the network's input channel from 1 to 3 in order to feed the original image inside. And changing the color_type parameter in LoadImageFromFile transform to color is required to prevent it converts any images into grayscale.
   

   mmocr/configs/textrecog/crnn/_base_crnn_mini-vgg.py

   Line 25 in f30c16c

   color_type='grayscale',

[htaoruan]

Thank you for your generous advice。