Image and Video Dehazing
An image may suffer from low contrast or poor resolution under hazy weather, and image dehazing is very important in image restoration. Current image dehazing techniques mainly consist of traditional approaches that utilise some fundamental relations or important observations and AI-based methods that depend on deep neural networks. In this lab report, we emphasised investigating traditional dehazing techniques. We benchmarked histogram equalisation, Dark Channel Prior, Single Scale Retinex, and Homomorphic filtering for image dehazing and Spatio-Temporal Markov Random Field (ST-MRF) for video dehazing on given datasets. While each method has its advantage in enhancing specific images, the Dark Channel Prior best qualitatively preserves unhazed images, and ST-MRF generated more natural results than the dehazed images at a faster speed per frame. Additionally, we proposed a novel method for generating dark channels that could restore the colour information better than the original Dark Channel Prior. Future works can focus on introducing quantitative metrics for dehazing performance evaluation.