Single-image high dynamic range reconstruction using attention based autoencoder

Single-image High Dynamic Range (HDR) reconstruction remains a significant challenge in modern image processing. In Low Dynamic Range (LDR) images, exposure limitations constitute a major obstacle to preventing information loss. In this thesis, a U-Net-like Encoder–Decoder based HDR generation framework was adopted, and the integration of different attention mechanisms into this structure was investigated. The proposed architecture was implemented within an autoencoder framework, where five attention mechanisms—Spatial Attention, Channel Attention, Bottleneck Attention, Squeeze-and-Excitation (SE), and Self-Attention—were incorporated as separate model variants. In addition, performance improvements were achieved by introducing different loss functions, most notably the Exposure Consistency Loss, into the training process. The developed models were trained on the DrTMO dataset and quantitatively evaluated using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity(LPIPS). The results demonstrate that the proposed attention mechanisms and additional loss functions significantly improve HDR reconstruction quality, with the Spatial Attention–based model achieving the highest performance in terms of both structural fidelity and perceptual quality. The findings are further supported by ablation studies and detailed visual comparisons.