FHDR: HDR Image Reconstruction from a Single LDR Image using Feedback Network

Abstract

High dynamic range (HDR) image generation from a single exposure low dynamic range (LDR) image has been made possible due to the recent advances in Deep Learning. Various feed-forward Convolutional Neural Networks (CNNs) have been proposed for learning LDR to HDR representations. To better utilize the power of CNNs, we exploit the idea of feedback, where the initial low level features are guided by the high level features using a hidden state of a Recurrent Neural Network. Unlike a single forward pass in a conventional feed-forward network, the reconstruction from LDR to HDR in a feedback network is learned over multiple iterations. This enables us to create a coarse-to-fine representation, leading to an improved reconstruction at every iteration. Various advantages over standard feed-forward networks include early reconstruction ability and better reconstruction quality with fewer network parameters. We design a dense feedback block and propose an end-to-end feedback network-FHDR for HDR image generation from a single exposure LDR image. Qualitative and quantitative evaluations show the superiority of our approach over the stateof-the-art methods.

Deep HDR estimation with generative detail reconstruction

“…An et al [ALK19] proposed an U-net like network architecture with partial convolutional layers in the encoder for inpainting the masked under-or over-exposed pixels in the input SDR image. Khan et al [KKR19] proposed an recurrent network architecture with feedback blocks for HDR reconstruction, which enables a coarse-to-fine representation at every iteration. Santos et al [STKK20] proposed an neural network which contains feature masking mechanism at each convolutional layer and fine-tuned on a pre-trained inpainting model for HDR reconstruction.…”

HDR-NeRF: High Dynamic Range Neural Radiance Fields

“…In contrast, alignment-based methods align the input multiple LDR images by estimating optical flow then merge the aligned images [25,49,53]. Depending on the great potential of deep learning, some methods try to reconstruct an HDR image from a single LDR image [15,26,27]. However, most HDR imaging methods require the given LDR images with a fixed or quasi-fixed camera pose.…”

Deep Learning for HDR Imaging: State-of-the-Art and Future Trends

“…Although the HDR image quality can be improved by increasing network depth or adding more losses, it involves considerable computational costs. FHDR, proposed by Khan et al [12], exploits…”

“…As such, low-level features are guided by high-level features in multiple iterations, leading to better reconstruction with few parameters. In contrast to [12], a different perspective of the feedback mechanism was proposed in [90]; the method involves learning to generate HDR images first. Next, the outputs are used to generate LDR images reciprocally via a correction network.…”

“…However, as capturing LDR contents at different exposures requires the use of certain software/hardware technology, it is often difficult to create datasets. Therefore, the second approach is to generate HDR content from a singleexposure image [10], [11], [12], [13], [14].…”