We developed a machine learning framework for automatically detecting temporal noise artifacts in video frames. The project focuses on identifying visually degraded regions at the level of 16×16 image patches, providing a practical tool for video quality assessment and analysis.

Unlike many recent approaches that rely on large deep learning models, this framework combines carefully engineered texture and frequency domain features with gradient-boosted decision trees as the classifier. By analyzing temporal differences between video frames, the system extracts complementary descriptors such as DCT coefficients, Haar wavelet energies, statistical measures, and local texture features. XGBoost-based feature selection and a lightweight linear projection are then used to retain the most informative characteristics before training the final XGBoost classifier.

To further improve prediction quality, the framework incorporates a spatial post-processing stage that refines neighboring patch predictions, producing smoother and more accurate distortion maps.

Designed with a modular and configurable architecture, this project provides an efficient and interpretable solution for video temporal noise detection while offering comprehensive visualization and evaluation tools.