PSHDR: HDR Image Reconstruction from a Single-Exposure LDR Using Residual Features

• Lucas Hildelbrano Costa Carvalho UFAC
• Quefren Leher UFAC
• Josue Lopez-Cabrejos UFAC
• Gustavo Ferreti UFAC
• Ana Beatriz Alvarez UFAC
• Thuanne Paixão UFAC
• Olacir Rodrigues Castro Junior UFAC

Resumo

Recent advances in deep learning techniques have led to the development of new approaches for reconstructing high dynamic range (HDR) images from a single low dynamic range (LDR) image. However, the images reconstructed by these methods show a loss of information in overexposed and underexposed regions. To overcome this challenge, this paper presents a deep neural network architecture designed to reconstruct HDR images from a single LDR image, with the aim of restoring information in saturated regions. The proposed architecture uses an encoder-decoder network that incorporates the Deformable Convolution Residual Block (DCRB) and Residual Feature Block (RFB) modules to restore information in saturated regions and a conditional network to modulate the information extracted from the input LDR image. To validate the proposed architecture, an experimental study was conducted using the NTIRE 2021 dataset, and the results were compared to 3 state-of-the-art algorithms. The results of the quantitative and qualitative analyses show that the proposed architecture obtains the best results among the compared approaches in the task of reconstructing singleexposure HDR images.

Palavras-chave: Deep learning, Visualization, Employment, Artificial neural networks, Feature extraction, Image restoration, High dynamic range, Data mining, Proposals, Image reconstruction