Introduction

Came across this amazing paper which aimed to improve the low light vision capability of the cameras that we use. Our own biological vision tries to figure out two key things when looking at an image:

• Instrisic Reflectance: Which is to explain the actual colour and texture of an object independent of the light conditions.
• Incident Illumination: What is the light source and how it is falling on the scene.

Retinex-Net: Trying to convert the above biological trickery into computer algorithm. This is where this deep learning algorithm plays an important role. Instead of a single network, Retinex-Net employs two interconnected subnetworks: Decomposition Network (Decom-Net) and Enhancement Network (Enhance-Net)

Working

Stage 1: Decomposition with Decom-Net This is the first stage of the Retinex-Net and it is designed to decompose the image into its reflectance and illumination maps.

Inputs:

• Pair of images;
  • S_low: Low light version of the scene.
  • S_normal: well lit version of the scene

Outpts:

• R_low, I_low: reflectance and illumination from the low light image
• R_normal, I_normal: reflectance and illumination from the normal light image

Network Architecture:

It is a FCN with:

1. 3x3 Conv Layer: Extracts the low level features of the image