Creation of DALT-NET: Deep Learning 3D Lookup Table Generation Approach to Daltonization for Dichromatic Vision

Color Vision Deficiency (CVD) is a condition which can inconvenience or even endanger 8% of men and 0.44% of women globally. To address this issue, this paper proposes the first deep learning and lookup-table based recoloring (Daltonization) method, termed Dalt-NET. DaltNET is composed of N learnable 3D Lookup Tables (LUT), one fixed failsafe 3D LUT, and a multi-task Convolutional Neural Network (CNN) to generate 3D LUT coefficients. This model employs an assembly architecture trained through enhanced data preprocessing. The advantages of the developed method are as follows: 1) Simulations show that Dalt-NET can achieve better CVD perceptual quality with more details, 2) it can transform an 8K image at 0.003s under a 0.3G FLOP MACs requirement, which is 3400 times faster and requires 30000 times less computations than previous methods, 3) it provides a novel framework for further development, such as model optimization, database augmentation, and specific spatiotemporal filtering. The proposed research opens a new door providing CVD individuals opportunities for the rapidly expanding media infrastructure and industry.

I. Introduction

Color sensitivity is a crucial means in the process of visual data collection from both digital media and physical needs. However, Color Vision Deficiency (CVD) affects 8% of men and 0.44% of women, denying 300 million [1] across the world both safety, comfort, and inherent opportunities available for color normal observers (CNO). Protanomaly, deuteranomaly, and tritanomaly are anomalous trichromacy, respectively issues in perceiving long, medium, and short wavelengths relative within the visible spectrum. Although there are different types and levels of CVD, red-green color blindness covers 95% of the people with CVD [1]. While protan blindness is used as an example during the illustration and algorithm explanations, the proposed algorithm can be applied to the other two types of CVD.