This project explores the effectiveness of various deep learning-based super-resolution (SR) models in improving both the visual quality and object detection performance of low-light images. It evaluates how perceptual quality (e.g., SSIM, PSNR) correlates with detection accuracy when used as preprocessing for YOLOv8.
- SRCNN – Classic 3-layer model for baseline performance.
- Enhanced SRCNN – Adds residual blocks for deeper feature extraction.
- Enhanced SRCNN v2 – Introduces Squeeze-and-Excitation (SE) channel attention.
- YASRNet – Combines spatial and channel attention via CBAM with PixelShuffle upsampling.
- SRCNN Edge – Uses a hybrid MSE + Laplacian loss to preserve edge details.
The super-resolved images were passed into a frozen YOLOv8 detector and compared against:
- Original low-light inputs
- High-light (ground truth) versions
- Super-resolved outputs
Metrics used:
- SSIM / PSNR for visual fidelity
- YOLOv8 Precision, Recall, and Exact Match Accuracy for detection performance
- LOLv1 Dataset for low-light image enhancement.
- Images are used at full resolution for preserving edge structures critical to detection.
- Framework: PyTorch
- GPUs: 2× NVIDIA T4 (16 GB VRAM)
- Mixed precision: Disabled
- Training Epochs: 30
- SR models improve perceptual quality, but not always detection performance.
- High SSIM does not imply high YOLO detection accuracy.
- Models that prioritize edge preservation (like SRCNN Edge) are more useful for downstream tasks.
