Real-ESRGAN is a PyTorch-based implementation of an advanced image enhancement model tailored to generate high-quality images from low-light inputs. The model integrates machine learning regressors with a lightweight Real-ESRGAN architecture to achieve superior results. Initial preprocessing steps, such as brightness and contrast adjustments, are performed by machine learning regressors before passing the image to the Real-ESRGAN for final enhancement. This hybrid approach demonstrates exceptional performance on the LOL dataset, achieving a remarkable Peak Signal-to-Noise Ratio (PSNR) of 25.9.
- Hybrid Architecture: Combines traditional machine learning regressors with Real-ESRGAN for effective low-light image enhancement.
- Preprocessing Pipeline: Applies basic image adjustments, such as brightness and contrast, through machine learning techniques.
- Lightweight Design: Optimized for performance while maintaining high-quality outputs.
- State-of-the-Art Performance: Outperforms existing models on the LOL dataset with a PSNR of 25.9.
The model consists of two primary components:
- Enhances the input image by adjusting parameters like brightness, contrast, and saturation.
- Acts as a preprocessing step to provide a better starting point for the deep learning model.
- A lightweight super-resolution GAN optimized for low-light image enhancement.
- Generates high-quality enhanced images by learning from the preprocessed outputs of the regressor.
The model is trained and evaluated on the LOL (Low-Light) dataset, which includes:
- Low-light input images.
- Corresponding high-quality reference images.
- PSNR: 25.9 (on the LOL dataset).
- SSIM: 0.81 (on the LOL dataset).
- Qualitative Results: Produces visually appealing enhanced images with well-preserved details and reduced noise.