Overcoming Degradation Imbalance for Consistent Image Dehazing

Pranjay Shyam (Faurecia IRYStec),* Hyunjin Yoo (Faurecia IRYStec)
The 34th British Machine Vision Conference


MSE or MAE loss functions work under the premise of all pixels having equal contribution during optimization. However, natural haze degradations are non-homogeneous, resulting in varied haze distribution. This results in sub-optimal training of current state-of-the-art (SoTA) supervised learning-based image dehazing algorithms due to an imbalance in pixel contribution. The outcome is the poor recovery of areas affected by severe degradations as these are underrepresented vis-a-vis mild and moderate affected regions during training. To address this data imbalance and generate consistent, visually pleasing restored images, we identify strategies at data augmentation and loss computation stages to ensure degradation-balanced training for image dehazing. From a data augmentation perspective, we propose a peak-signal-to-noise ratio (PSNR) based patch sampling mechanism and an adversarial auto-augmentation mechanism to vary the degradation distribution intensity within training samples. Second, to reduce the bias introduced by increasing the proportion of accurately recovered pixels along the training cycle, we propose focal pixel loss that scales the contribution of individual pixels towards loss calculation with restoration accuracy as prior. We successfully integrate the proposed focal loss in current pixel and feature-based loss functions. Finally, to ensure perceptually pleasant and structurally accurate image restoration, we propose a dynamic version of contrastive regularization with dynamic boundary constraints to better constrain the latent representations. We demonstrate the proposed mechanisms' efficacy in improving the performance of SoTA image dehazing algorithms without modifying the underlying network architecture.


author    = {Pranjay Shyam and Hyunjin Yoo},
title     = {Overcoming Degradation Imbalance for Consistent Image Dehazing},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year      = {2023},
url       = {https://papers.bmvc2023.org/0601.pdf}

Copyright © 2023 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection