Gradient-based metric for image defogging without ground truth

Gerard deMas-Giménez, Pablo García-Gómez, Josep R. Casas and Santiago Royo

Abstract

Fog, haze, or smoke are usual atmospheric phenomena that dramatically compromise the overall visibility of any scene, critically affecting features such as illumination, contrast, and contour detection of objects. The decrease in visibility compromises the performance of computer vision algorithms such as pattern recognition and segmentation, some of them very relevant for decision-making in the fields of security and autonomous vehicles. Several dehazing methods have been proposed. However, to the best of our knowledge, all proposed metrics in the literature compare the defogged image to its ground truth for evaluation of the defogging algorithms, or need to estimate parameters through physical models. This fact hinders progress in the field as obtaining proper ground truth images is costly and time-consuming, and physical parameters greatly depend on the scene conditions. This paper aims to tackle this issue by proposing a gradient-based metric for image defogging evaluation that does not need a ground truth image or a physical model. The proposed metric only requires the original hazy RGB image and the RGB image obtained after the defogging procedure. A comparison of the proposed metric with metrics already used in the NTIRE 2018 defogging challenge is performed to prove its effectiveness in a general situation, showing comparable results to conventional metrics.

In this repository you will find the MATLAB implementation of the Contour-based metric for image defogging without ground truth algorithm found in the paper.

Edge comparison between a fogged image (left) and its fog-free ground truth (right). Both colour images are presented on top with their associated edge images below. Both images are part of the O-Haze dataset.

Usage

Contours, edges, and textures are greatly affected by fog. Several proposes have been made to defog an image, including the use of end-to-end Deep Neural Networks. In order to quantify the enhancement of the defogging method, we perform the derivative of both, the original hazy image and the proposed defogged one. After that, we compute the pixel-by-pixel relative difference between the two differentiated images. Then, we perform the histogram of that relative difference matrix. In order to obtain a quantitative value that measures the defogging enhancement we compute the weightned ratio between the positive and negative part of the histogram, that is the final $R$ value. $R$ can take values from -1 to 1, being 1 when all the gradients have been enhanced and -1 when the defogging procedure have worsened every gradient of the image. The weighted character of the metric is used to strengthen the gradient that have been greatly improved or worsened. In the image above, our metric's value is $R=0.9844$ which is consistent with the fact that the defogged image is, in fact, the actual Ground Truth of the scene.

The algorithm is implemented in the Contour_based_metric_woGT.m on the Demo_dataset. Contour_based_metric_woGT.m will output an image as showed above. The images on the first row are the Hazy and Defogged images. On the second row we can find the relative difference histogram, with the $R$ value of our metric, as well as the relative difference image. White points correspond to the background, green shows us the points where a gradient is enhanced, red, the contrary. The intensity of the difference is conserved, so the brighter color the higher the relative difference between the gradient of the Hazy image compared to the Defogged one. The functions used in the file can be found at the Functions_metric folder. Please, if our metric is useful for your work, consider citing our paper.

DeMas-Giménez, G.; García-Gómez, P.; Casas, J.R.; Royo, S. Contour-Based Metric for the Evaluation of Image Defogging without a Ground Truth. Preprints 2022, 2022090127 (doi: 10.20944/preprints202209.0127.v1).

Demo dataset

Apart from the implemented algorithm, the user may find a demo dataset as well. This dataset is part of the O-Haze dataset used in the NTIRE 2018 defogging challenge. If you use the images for academic porpuses make sure to cite the authors properly.

@inproceedings{O-HAZE_2018,
author = { Codruta O. Ancuti and Cosmin Ancuti and Radu Timofte and Christophe De Vleeschouwer},
title = {O-HAZE: a dehazing benchmark with real hazy and haze-free outdoor images},
booktitle =  {IEEE Conference on Computer Vision and Pattern Recognition, NTIRE Workshop },
series = {NTIRE CVPR'18},
year = {2018},
location = {Salt Lake City, Utah, USA},
}