Q-DRIR: A Quaternion-Based Framework for Detail Restoration and Illumination-Robust Exposure Correction

Images captured by consumer-grade devices often exhibit underexposure or overexposure due to the limited dynamic range of image sensors under complex lighting conditions, resulting in significant loss of visual details. To address this issue, we propose a novel a quaternion-based framework for detail restoration and illumination-robust exposure correction (Q-DRIR). Q-DRIR leverages a pre-trained DDPM as a feature prior, extracting its multi-scale and multi-timestep intermediate features. These features are jointly fed into an upsampling decoder along with pyramid features generated by a downsampling encoder from the original image, progressively reconstructing a high-quality exposure-corrected image. The upsampling decoder comprises two core components: (1) the Quaternion Diffusion Detail Enhancement Module (QDDEM), which employs Quaternion Multi-Sparsity Attention (Q-MSA) to adaptively focus on critical features while suppressing noise interference; and (2) the Quaternion Illumination Adaptation Module (QIAM), which utilizes Quaternion Differentiable Local Sensitive Hashing (QD-LSH) to achieve illumination-aware soft quantization and propagates contextual information via bidirectional recurrent filtering, enhancing robustness to complex lighting conditions. To the best of our knowledge, extensive experiments on multiple benchmark datasets of low-light and overexposed images demonstrate that Q-DRIR outperforms State-Of-The-Art (SOTA) methods in underexposure correction and achieves significant improvements in enhancing overexposed images, showcasing its strong generalization capability and robustness in real-world electronic imaging applications.