This project focuses on the Low-Light Image Processing Transformer, a neural network architecture designed to enhance images captured in low-light conditions. The study focuses on fine-tuning the hyperparameters of an existing model (HVI-CIDNet - https://github.com/Fediory/HVI-CIDNet) designed to improve perceptual metrics. This is accomplished by using a stable diffusion approach to adjust the lighting of a source image instead of adding noise. This innovative model processes brightness and color features separately using a dual-path architecture. The project also introduces Fréchet Inception Distance (FID) as a new evaluation metric for image quality, alongside conventional metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measurement (SSIM).

1. Dual-Path Architecture:

   • Separates processing of brightness and color features.
   • Employs the Horizontal/Vertical-Intensity (HVI) transformation for efficient color space conversion.

2. CIDNet:

   • A UNet-based architecture featuring two branches for brightness and color processing.
   • Uses the Lightweight Cross-Attention (LCA) module for information exchange and noise suppression.

3. Evaluation Metrics:

   • FID to assess generated image similarity to real images.
   • PSNR and SSIM for traditional quality assessment.

4. Dataset:

   • A curated dataset from Kaggle consisting of 500 paired low-light and normal-light images.

Source: Kaggle [https://www.kaggle.com/datasets/soumikrakshit/lol-dataset]
Content: 500 paired images (low-light and normal-light), predominantly indoor scenes with a resolution of 400×600.
Division: 485 training pairs, 15 testing pairs.

1. Horizontal/Vertical-Intensity (HVI) Transformation
    * Intensity Map: Derives the maximum value from RGB channels.
    * HV Color Map: Measures color reflectance with adjustable density.
    * Trainable HVI Map: Combines intensity and HV color maps for reversible transformation.

2. Color and Intensity Decoupling Network (CIDNet)
    * HV-way: Focuses on color information processing.
    * Intensity-way: Handles brightness adjustments for varied lighting scenarios.
    * LCA Module:
        * Includes Cross-Attention Block (CAB), Intensity Enhance Layer (IEL), and Color Denoise Layer (CDL).

3. Fréchet Inception Distance (FID)
    * Compares feature distributions of generated and real images.
    * Lower FID indicates higher quality.

Three model variants were evaluated.
Model 3 performed best with optimal metrics:
    PSNR: 24.54 dB
    SSIM: 0.8633
    FID: 31.07

The Low-Light Image Processing Transformer demonstrates significant improvement in image enhancement tasks under low-light conditions. While small adjustments can be time-consuming, increasing model complexity improves performance. Future goals include integrating differentiable FID in the loss function and leveraging advanced techniques like LiDAR for improved accuracy and precision.

1. Dataset: Kaggle LOL Dataset [https://www.kaggle.com/datasets/soumikrakshit/lol-dataset]
2. HVI-CIDNet: You Only Need One Color Space: An Efficient Network for Low-light Image Enhancement [https://github.com/Fediory/HVI-CIDNet]
3. pytorch-frechet-inception-distance: Fréchet Inception Distance (FID) for Pytorch [https://github.com/hukkelas/pytorch-frechet-inception-distance]