An Efficient and Low-Complexity Transformer-Based Deep Learning Framework for High-Dynamic-Range Image Reconstruction

High-dynamic-range (HDR) image reconstruction involves creating an HDR image from multiple low-dynamic-range images as input, providing a computational solution to enhance image quality. This task presents several challenges, such as frame misalignment, overexposure, and motion, which are addressed using deep learning algorithms. In this context, various architectures with different approaches exist, such as convolutional neural networks, diffusion networks, generative adversarial networks, and Transformer-based architectures, with the latter offering the best quality but at a high computational cost. This paper proposes an HDR reconstruction architecture using a Transformer-based approach to achieve results competitive with the state of the art while reducing computational cost. The number of self-attention blocks was reduced for feature refinement. To prevent quality degradation, a Convolutional Block Attention Module was added, enhancing image features by using the central frame as a reference. The proposed architecture was evaluated on two datasets, achieving the best results on Tel's dataset in terms of quality metrics. The computational cost indicated that the architecture was significantly more efficient than other Transformer-based approaches for reconstruction. The results of this research suggest that low-complexity Transformer-based architectures have great potential, with applications extending beyond HDR reconstruction to other domains.