LIGHTWEIGHT ADAPTIVE PRE-PROCESSING FOR ROBUST FACE RECOGNITION IN LOW-LIGHT CONDITIONS

Abstrak

Face recognition works great when the lighting is good, but once things get dark, performance drops fast. This project looks at a simple fix — instead of retraining complicated models or using special hardware, I designed a lightweight pre-processing step that cleans up low-light images so face recognition systems can handle them better. The module uses a basic U-Net setup and learns to improve image quality while keeping the important details that define someone’s face. I tested the system on low-light images created from public datasets, and it showed a clear improvement in recognition, boosting identity similarity by over 36%. It also runs fast enough for real-time use, even on average hardware.