FLASH: Fast and Lightweight Architecture Using State Space Models for HDR Multi-Exposure Reconstruction

Multi-exposure high dynamic range (HDR) image reconstruction makes it possible to improve the quality of low dynamic range (LDR) images using algorithms based on deep learning. The main challenge in this task lies in the generation of ghosting artifacts in the resulting image, caused by the displacement of objects and by the low or overexposure in the input LDR images. To overcome this challenge, this paper proposes the FLASH-S architecture, which incorporates an innovative attention block using Vision Mamba state-space models (SSM) to enhance the extracted features from the input images. The features are serialized in four different directions to guarantee the same importance in each patch of the sequence in the SSM. In addition, the FLASH architecture was proposed, an improved version of FLASH-S using optical flow to align the input images. Experiments were carried out using the Kalantari, SynHDR and Tursun datasets. Based on seven quality metrics and three computational cost metrics, the results showed that the FLASH-S and FLASH proposals are superior to state-of-the-art algorithms. Additionally, FLASH demonstrated better reconstruction quality in qualitative evaluations, generalization capacity and tests in real environments. The model’s source code can be obtained from: https://github.com/pavic-lab-ufac/FLASH.