PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing
Authors: Huaizhuo Liu, Hai-Miao Hu, Yonglong Jiang, Yurui Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate the superior performance of our PEIE method, significantly surpassing the state-of-the-arts in real-world dehazing. We compared our method with state-of-the-art methods through comprehensive experiments. Both subjective evaluations and no-reference quality metrics demonstrate that PEIE achieves superior dehazing performance and robust adaptation to diverse illumination conditions. All experiments are implemented using the Py Torch framework on 4 NVIDIA RTX 2080Ti GPUs. |
| Researcher Affiliation | Academia | 1Hangzhou Innovation Institute, Beihang University 2State Key Laboratory of Virtual Reality Technology and Systems, Beihang University 3School of Computer Science and Engineering, Beihang University EMAIL |
| Pseudocode | No | The paper describes the proposed method in descriptive text, mathematical equations (e.g., Eq. 1, 2, 3), and architectural diagrams (e.g., Figure 4), but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | More details can be found at https://github.com/Hankitle/PEIE. |
| Open Datasets | Yes | We first assess the visual quality of PEIE on real hazy images from the RTTS subset of the RESIDE dataset (Li et al. 2019), which contains 4,322 real-world hazy images. We utilize 500 clean images from RIDCP (Wu et al. 2023) as input in our data synthesis pipeline, guiding network training through online synthesis. To further demonstrate PEIE s superior generalization ability, we also evaluate it on Fattal s dataset (Fattal 2015), which includes 31 classic real hazy cases. We conducted a human subjective study to compare PEIE s performance with other methods using 10 real-world hazy images from the HSTS subset of RESIDE. |
| Dataset Splits | No | We utilize 500 clean images from RIDCP (Wu et al. 2023) as input in our data synthesis pipeline, guiding network training through online synthesis. Our proposed IADNet is trained on data generated by the synthesis pipeline over 10,000 iterations. The paper mentions datasets used for evaluation (RTTS, Fattal's, HSTS subsets of RESIDE) but does not provide specific training/validation/test splits or methodologies for the data used during training or evaluation. |
| Hardware Specification | Yes | All experiments are implemented using the Py Torch framework on 4 NVIDIA RTX 2080Ti GPUs. |
| Software Dependencies | No | All experiments are implemented using the Py Torch framework on 4 NVIDIA RTX 2080Ti GPUs. The paper mentions the PyTorch framework but does not specify its version or other software dependencies with version numbers, which are needed for full reproducibility. |
| Experiment Setup | Yes | Training is conducted using the Adam optimizer with default parameters (β1 = 0.9, β2 = 0.99), a fixed learning rate of 1 x 10^-4, and a batch size of 16. Data augmentation includes random resizing and cropping to 256 x 256, flipping with a 50 percent probability, and the addition of noise. Our proposed IADNet is trained on data generated by the synthesis pipeline over 10,000 iterations. We utilize L1 loss, perceptual loss, and GAN loss which are widely used in image restoration as objective functions, with weights of 1.0, 1.0, and 0.1, respectively. In the IDU, we employ average pooling layers with sizes [4, 8, 12, 16, 24], where the size scales with the feature map size. For A, we randomly generate values within the range [0.8, 1.0]. For transmission, we randomly generating β [0.8, 2] to control haze density. |