Translation of Low Illumination On-road Images to Well Illuminated Ones for Assisting Driving using Deep Auto-Encoders

Low illumination can be dangerous for driving, as human perception depends upon ambient illumination. This paper attempts to aid driving in poor illumination using illumination-correction techniques. We propose an approach of low-light image enhancement using a de-noising convolutional auto-encoder (DNCAE) dedicated to on-road conditions. We prepare a custom data set consisting of paired images of well-lit and synthetically darkened images. A sample of this data set is used to train a neural network model that produces a better-illuminated estimate of the scene. We solve the color constancy issue of our model using a color space conversion designed for human perception of vision. The method has shown effectiveness against existing illumination correction techniques regardlna specific metrics viz. PSNR, SSIM, MSE and MAE.