Image dehazing with machine learning

Abstract

Images can be degraded due to various reasons. Degraded images suffer from low visibility, loss of contrast, color distortion, etc. To restore the visual quality, image dehazing can be applied. This project aims to use the Machine Learning method to perform image dehazing. The machine learning method adopted in the project is called AOD-Net (All-in-One Dehazing Network) which is a lightweight Convolution Neural Network. Moreover, this project has done various modifications to the AOD-Net to improve the quality of the dehazed image. The convolution layers and the activation function has been revolutionized. The blank model has been trained a lot of times using the NYU2 dataset to optimize the results. The model is evaluated on synthetic and real hazy images with quantitative metrics. The testing dataset are FRIDA2 and BeDDE and there are three metrics used to evaluate the result which are SSIM, Visibility Index and Realness Index. We also compare AOD-Net and modified version with other image dehazing methods such as Dark Channel Prior(DCP). The modified AOD-Net has outperformed other methods compared in this project.