AGNet: Attention guided network for single HDR reconstruction

Single-image HDR reconstruction aims to recover an HDR image from a single LDR input. However, under extreme degradations such as severe noise and large overexposed regions, existing methods often suffer from color distortion and detail loss. To address these challenges, we propose AGNet, an attention-guided end-to-end network with three key advantages. First, to effectively handle large overexposed regions, we introduce a global spatial attention mechanism to enable long-range semantic modeling at a relatively low computational cost, endowing the model with the capability to leverage global contextual information for restoring saturated areas. Second, to mitigate color distortion caused by overfitting in LDR-to-HDR mapping, we design a lightweight branch that constrains the mapping complexity. Third, to suppress noise during dequantization, we incorporate a gradient-guided channel attention module, which utilizes gradient priors to adaptively suppress noise. Experimental results demonstrate that AGNet significantly outperforms existing methods in single-image HDR reconstruction, achieving superior reconstruction quality while maintaining low parameter complexity. Our code is available at https://github.com/gzzzzzz781/AGNet.