PavicHDR: An Efficient Architecture for HDR Image Reconstruction Based on CNN and Attention Mechanisms

• Josue Lopez-Cabrejos UFAC
• Quefren Leher UFAC
• Gustavo Ferreti UFAC
• Lucas Hildelbrano Costa Carvalho UFAC
• Thuanne Paixão UFAC
• Ana Beatriz Alvarez UFAC
• D. B. Luque UFAC

Resumo

High dynamic range (HDR) image reconstruction is a technique that consists of merging multiple low dynamic range (LDR) images of the same scene in order to increase the quality of the image. One of the main challenges in this process is the movement present in each frame, which can generate ghosting artifacts in the resulting image. This paper proposes the PavicHDR architecture, designed to eliminate ghosting due to movement while maintaining a low computational cost. To do this, efficient feature extraction, alignment, refinement and fusion techniques were used. The quantitative and qualitative experiments carried out with the Kalantari and Tel datasets show the superior performance of PavicHDR in terms of reconstruction quality, reduction of distortions and preservation of structures. Moreover, it stands out for its reduced inference time compared to the state-of-the-art, including methods based on diffusion, vision transformers and adversarial neural networks.

Palavras-chave: Graphics, Computer vision, Neural networks, Merging, Computer architecture, Transformers, Feature extraction, Distortion, High dynamic range, Image reconstruction