LightClearNet: Lightweight Encoder-Decoder Architecture for Single Image Dehazing

Image dehazing is a crucial low-level vision task that recovers clear images from hazy ones. In recent years, immense progress has been made in recovering the original image as accurately as possible, without any consideration for efficiency and performance on low-end hardware, making it a mere relic for image beautification. Image dehazing could be very critical for enhanced visibility during indoor emergencies like fire hazards, utilization in surveillance and security cameras and in autonomous vehicles, and hence should be quick and easy to compute in order to be accessible. Therefore, we propose LightClearNet, a lightweight yet just as effective model for single image dehazing based around batch normalization and skip connections. The model has been trained on RESIDE-6K dataset and our lightweight model delivers comparable results to DehazeFormer while being twice as performant as the lightest variant of DehazeFormer.