Learning to Intrinsic Image Filter for Instagram Filter Removal
Seohyeon Lee, Guisik Kim, Junseok Kown · 2022 13th International Conference on Information and Communication Technology Convergence (ICTC) · 2022
The filter removal task is important because filtered images risks degrading the performance of the computer vision model. We propose a two-branch model which performs filter removal task. Our two-branch model consists of a Palette based Un-filtering Model (PUM) and a Palette Injection Model (PIM). PUM learns an intrinsic filter from the input image using the color palette. It is simple and fast to remove the filter from the input with the learned intrinsic filter. PIM has VGG baseline model as an encoder, and injects palette on each decoding stage. The output is an unfiltered image itself. Our method fuses these two unfiltered results and obtains the final result. This ensures that the filter is removed accurately and the structure of image is maintained. As a result of the experiment, our proposed model performs better on filter removal task than other recent models.