I-HAZE: a dehazing benchmark with real hazy and haze-free indoor images (1804.05091v1)

Abstract: Image dehazing has become an important computational imaging topic in the recent years. However, due to the lack of ground truth images, the comparison of dehazing methods is not straightforward, nor objective. To overcome this issue we introduce a new dataset -named I-HAZE- that contains 35 image pairs of hazy and corresponding haze-free (ground-truth) indoor images. Different from most of the existing dehazing databases, hazy images have been generated using real haze produced by a professional haze machine. For easy color calibration and improved assessment of dehazing algorithms, each scene include a MacBeth color checker. Moreover, since the images are captured in a controlled environment, both haze-free and hazy images are captured under the same illumination conditions. This represents an important advantage of the I-HAZE dataset that allows us to objectively compare the existing image dehazing techniques using traditional image quality metrics such as PSNR and SSIM.

• The paper introduces the I-HAZE dataset featuring 35 real paired indoor images captured using a professional haze machine for authentic atmospheric replication.
• It employs objective evaluation with standard metrics like PSNR, SSIM, and CIEDE2000 to compare various single-image dehazing methods.
• The experimental findings highlight inconsistent algorithm performance, underscoring challenges in accurate dehazing and the need for enhanced restoration models.

An Evaluation of I-HAZE: A Comprehensive Dehazing Benchmark

The paper "I-HAZE: a dehazing benchmark with real hazy and haze-free indoor images" introduces an innovative dataset aimed at objectively assessing and comparing image dehazing methods. The scarcity of ground-truth images for effective benchmarking poses significant challenges to the dehazing domain. Thus, I-HAZE emerges as a crucial contribution, comprising 35 pairs of real hazy and corresponding haze-free (ground-truth) indoor images. This dataset extends beyond existing databases that often rely on synthetic haze, generated through computational models.

A distinguishing aspect of I-HAZE is the creation of hazy images using a professional haze machine, ensuring the replication of authentic atmospheric conditions. Included within each captured scene is a MacBeth color checker, which aids in color calibration and enhances the evaluation framework for dehazing algorithms. Critically, both hazy and haze-free images in the dataset are taken under identical lighting conditions, establishing a robust platform for deploying traditional image quality metrics like PSNR and SSIM for quantitative evaluations.

The paper also incorporates a comprehensive evaluation of several renowned single-image dehazing methods. It highlights the challenges inherent to single image dehazing, given the mathematically ill-posed nature of the task—primarily due to its dependence on variable transmission coefficients across image pixels.

Notably, the experimental results reveal crucial insights into the performance of existing dehazing algorithms. Despite the advanced methodologies employed, the benchmark suggests a general inadequacy among many tested techniques in accurately reconstructing original scenes from their hazy representations. The SSIM and CIEDE2000 metrics confirm room for further methodological advances in the field.

The comparisons reveal that the techniques proposed by Berman et al., Ancuti et al., and Ren et al. manifest superior average results when considering combined metrics of SSIM, PSNR, and CIEDE2000. Nonetheless, no universal framework emerges as the best performer across all images, indicating the task's complexity. The performance divergence underscores prevalent structural distortions and color corrections in dehazed images, rooted in transmission map and airlight estimation inaccuracies.

The implications of this paper extend beyond its immediate contributions. I-HAZE establishes vital groundwork for objectively enhancing dehazing algorithms, inviting further research to refine existing models or explore novel approaches capable of overcoming the documented limitations. As machine learning continues to integrate into image restoration domains, datasets like I-HAZE will be instrumental in driving more sophisticated, learning-based dehazing methodologies. The formal release of I-HAZE promises to be a stepping stone toward this progressive aim, heralding future advancements in image processing tasks where environmental degradation is a considerable hindrance.