Image Dehazing Using a Simple Convolutional Autoencoder
Percy Maldonado-Quispe, Hélio Pedrini · 2023
In this work, we introduce an end-to-end autoencoder-based convolutional neural network designed to effectively remove haze from hazy images. Our proposed approach leverages a supervised learning methodology, utilizing a dataset that encompasses both haze and haze-free images. Unlike traditional methods relying on atmospheric scattering models, our method aims to directly capture the intricate relationship between hazy and haze-free images. To train our model, we used the RESIDE dataset, consisting of diverse indoor and outdoor images, all with a size of 512$\times$ 512 pixels. Although our technique stands out for its simplicity and lightweight nature, it achieves competitive effectiveness according to evaluation metrics, properly establishing the correlation between haze and haze-free images.