A Deformable Convolutional Neural Network with Dual Attention for Multi-Exposure HDR Reconstruction

• Quefren Leher UFAC
• Josue Lopez-Cabrejos UFAC
• Gustavo Ferreti UFAC
• Lucas Hildelbrano Costa Carvalho UFAC
• Ana Gabriela Alvarez UFAC
• Thuanne Paixão UFAC
• Ana Beatriz Alvarez UFAC
• Facundo Palomino-Quispe UNSAAC

Resumo

The reconstruction of High Dynamic Range (HDR) images from multiple Low Dynamic Range (LDR) images with varying exposures allows for the recovery of details in regions of a scene that are either very bright or very dark. However, factors such as movement, camera instability, and sudden changes in lighting can cause misalignment and visual artifacts in the reconstructed HDR image, representing important challenges in the field. In this context, this paper presents a novel model for multi-exposure HDR reconstruction that integrates specialized sub-networks for feature extraction and enhancement, alignment, and fusion. Specifically, spatial and channel attention mechanisms, deformable convolutions and cross-attention are used to extract, align and combine features from LDR images. The effectiveness of the proposed model was demonstrated through qualitative and quantitative evaluations on the Kalantari and Tel HDR datasets, and the results were compared other state-of-the-art methods. The results demonstrated that the model effectively preserved spatial consistency in the reconstructed images, achieving superior performance on the Kalantari dataset and competitive results on the Tel dataset.

Palavras-chave: Deformable models, Deep learning, Visualization, Computational modeling, Lighting, Feature extraction, Computational efficiency, High dynamic range, Convolutional neural networks, Image reconstruction