DEEPSAFE
BRIEF
Synthetic facial media requires structured preprocessing pipelines that can isolate facial regions consistently before feeding frames into convolutional classification architectures.
ENGINEERING IDEA★
Build a modular Python pipeline combining OpenCV face localization and cropping with a TensorFlow CNN training architecture, paired with an interactive Streamlit prototype interface.
TOOLS USED
Python · OpenCV · TensorFlow · Streamlit · NumPy
PROJECT TYPE
Media Ingestion & Frame Decoding
Load input imagery and convert color spaces into normalized NumPy arrays.
OpenCV Face Localization & Cropping
Detect facial bounding boxes, extract aligned face regions, and resize tensors for CNN input.
TensorFlow CNN Training Pipeline
Convolutional feature extraction layers and binary classification head coded in TensorFlow.
Streamlit Demonstration Interface
Interactive prototype UI illustrating the user flow (currently operating with simulated demonstration predictions).
- ★Programmed OpenCV face-detection and image-preprocessing routines to isolate facial crops from raw input images.
- ★Authored TensorFlow convolutional neural network architecture and training pipeline scripts for real-versus-synthetic classification experiments.
- ★Built a Streamlit interactive prototype interface to demonstrate the end-to-end inspection workflow.
- ★Documented the current engineering state transparently: separating the implemented training/preprocessing code from the Streamlit demonstration layer's simulated predictions.
- ✓Established a working computer vision preprocessing and CNN training codebase in Python, OpenCV, and TensorFlow.
- ✓Created an interactive Streamlit demonstration interface for presenting the deepfake inspection workflow concept.
- →Bridging the gap between offline CNN training scripts and production inference requires serializing validated model checkpoints and benchmarking generalization across unseen compression artifacts.
- →Next engineering milestone is replacing the Streamlit demo's simulated prediction step with live inference against a validated model checkpoint.