A modified Dual Hierarchical Network to remove Shadows from Documents with Reflective Surface and Textured Backgrounds such as Sri Lankan National Identity Card (NIC)
This is a modified DHAN network with alteration to its base architecture based on Ghost Free Shadow Removal by Vinthony (https://github.com/vinthony/ghost-free-shadow-removal)
- DHAN is designed to tackle natural image shadows.
- It does a transfer learning using object detection / feature map (VGG19) in its pipeline to remove shadows.
- However, VGG19 is developed with the intent of classification of image into some objects (such as face, hair etc.)
- Our NIC dataset has neither of these features. Thus, creating a feature map to our image will result in wash out of useful textures and letters.
- Since VGG-19 is a CNN with 19 layers deep, it has difficulties working with high resolution images
- However practical photos taken with smartphones can produce high resolution images
- Instead of sending source image via VGG-19 and then into DHAN, we directly send source image into DHAN.
- Then to teach the GAN network to omit background textures, we send both source as well as taget images into discriminator and create a perception loss which the generator can use to adjust its next iteration of learning.
- We are in process of publishing our paper with regards to this project
