From 94f8032a63753bd01e5a8be9dcd74606a6512b3d Mon Sep 17 00:00:00 2001 From: Andras Schmelczer Date: Sun, 30 Jun 2024 22:22:16 +0100 Subject: [PATCH] Update README.md --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index a9b0bf5..c73b290 100644 --- a/README.md +++ b/README.md @@ -4,7 +4,7 @@ > Example of the network enhancing the colour of old digital photographs. -This project trains a neural network for automatically editing the style of digital photographs by learning a mapping from histograms of "bad" images to their aesthetic counterparts. Thus, both the inputs and outputs of the network are 3D RGB histograms: $$\text{bin}_{\text{red}} \times \text{bin}_{\text{green}} \times \text{bin}_{\text{blue}} \to \text{bin}'_{\text{red}} \times \text{bin}'_{\text{green}} \times \text{bin}'_{\text{blue}}$$ +This project trains a neural network for automatically editing the style of digital photographs by learning a mapping from histograms of "bad" images to their aesthetic counterparts. Thus, both the inputs and outputs of the network are 3D RGB histograms: $`\text{bin}_{\text{red}} \times \text{bin}_{\text{green}} \times \text{bin}_{\text{blue}} \to \text{bin}'_{\text{red}} \times \text{bin}'_{\text{green}} \times \text{bin}'_{\text{blue}}`$ By only exposing histograms to the network, we allow it to learn style-tranfer while eliminating the risk of changing the underlying structure of the source image in the process which is a shortcoming of existing deep learning-based approaches [^1] & [^2]. At the same time, non-linear transformations of the RGB colour distribution allow for much greater flexibility than predefined global adjustment tools such as _Brightness_ or _Contrast_.