Dalt-Net is the older model which set up a basis for InnoColor. As opposed to traditional methodology for Daltonization, which involves strict, determinate calculations that decide the rotation of color or color clusters within an image, Dalt-Net involves a hybrid deep-learning architecture. This is also different from the GAN-based deep learning method.
In terms of objectives met—InnoColor surpasses all previous methods, save the distillation data source, in cross-the-board performance. Check some of InnoColor's result comparisons here
Here are several issues in current algorithms, and how Dalt-NET resolves them.
1. Traditional approaches often follow a single scheme
However, what scheme recolors an image “best” (based on Daltonization standards) is heavily dependent on image color tones and lighting conditions. This is based on how the human brain perceives vision (which is interpreted mainly through the opponent color theory). So one method cannot work for all images.
2. Both traditional and GAN frameworks perform slowly
Traditional methods require searching for color clusters, which has a time complexity scaling based on image size (or volume for videos). Additionally, evaluation of GAN requires recursive calculations based on each pixel, and no explicit transformation can be generalized to all pixels which would increase speed.
1. Dalt-NET represents high dimensional and generalized data through 3D Lookup Tables
Dalt-NET is able to represent distinct and separable methods through its custom dataset design, which has been partitioned into a traditional (hue-rotation-based) data for baseline accuracy and novel (usually specialized and context-based recoloring) data. Additionally, it generates multiple 3DLUTs (3D Lookup Tables) to create adaptive channels based on the data, while maintaining a fixed 3DLUT as a performance failsafe.
2. Dalt-NET evaluates model based on downsampled image
Dalt-NET's weight predictor model is a simple, lightweight convolutional network whose goal is to simply process the color tones and lighting conditions of the original image. Specific object detection is not necessary. Thus, it can operate on low-resolution samples. This allows the convolutional network to output at near-constant time complexity, without hindering accuracy. In addition, the use of 3DLUTs avoids heavy algorithm-specific calculations by the computer.
Finally, check out the applications on this page! Click here
InnoColor utilizes a deep-learning guided image transformation to modify visual content, like images, video, and GUI, in both digital software and physical interfaces.
