Frequency-domain loss function for deep exposure correction of dark images

AbstractWe address the problem of exposure correction of dark, blurry and noisy images captured in low-light conditions in the wild. Classical image-denoising filters work well in the frequency space but are constrained by several factors such as the correct choice of thresholds and frequency estimates. On the other hand, traditional deep networks are trained end to end in the RGB space by formulating this task as an image translation problem. However, that is done without any explicit constraints on the inherent noise of the dark images and thus produces noisy and blurry outputs. To this end, we propose a DCT/FFT-based multi-scale loss function, which when combined with traditional losses, trains a network to translate the important features for visually pleasing output. Our loss function is end to end differentiable, scale-agnostic and generic; i.e., it can be applied to both RAW and JPEG images in most existing frameworks without additional overhead. Using this loss function, we report significant improvements over the state of the art using quantitative metrics and subjective tests.

Download Full-text

Single Image Dehazing Using End-to-End Deep-Dehaze Network

Electronics ◽

10.3390/electronics10070817 ◽

2021 ◽

Vol 10 (7) ◽

pp. 817

Author(s):

Masud An-Nur Islam Fahim ◽

Ho Yub Jung

Keyword(s):

Feature Extraction ◽

Loss Function ◽

Traditional Approach ◽

Real Life ◽

Weather Conditions ◽

Image Dehazing ◽

Image Translation ◽

End To End ◽

Nonlinear Feature Extraction ◽

Nonlinear Feature

Haze is a natural distortion to the real-life images due to the specific weather conditions. This distortion limits the perceptual fidelity, as well as information integrity, of a given image. Image dehazing for the observed images is a complicated task because of its ill-posed nature. This study offers the Deep-Dehaze network to retrieve haze-free images. Given an input, the proposed architecture uses four feature extraction modules to perform nonlinear feature extraction. We improvise the traditional U-Net architecture and the residual network to design our architecture. We also introduce the l1 spatial-edge loss function that enables our system to achieve better performance than that for the typical l1 and l2 loss function. Unlike other learning-based approaches, our network does not use any fusion connection for image dehazing. By training the image translation and dehazing network in an end-to-end manner, we can obtain better effects of both image translation and dehazing. Experimental results on synthetic and real-world images demonstrate that our model performs favorably against the state-of-the-art dehazing algorithms. We trained our network in an end-to-end manner and validated it on natural and synthetic hazy datasets. Our method shows favorable results on these datasets without any post-processing in contrast to the traditional approach.

Download Full-text