Deep Convolutional Neural Network (DCNN) Embedding
Definition
The face descriptor produced by a neural network trained to map face images into a metric space where same-person images cluster together; the foundation of all modern operational face recognition systems.
- Function
- Maps a face image to a numeric vector in a metric space
- Clustering property
- Same-person images produce nearby vectors
- Use
- Foundation of modern operational face recognition systems
- Comparison method
- Distance or similarity score between two embeddings
- Field
- Forensic facial comparison, NIST FRVT, FBI NGI
Common questions
How does a DCNN embedding differ from older feature-based face recognition methods?+
Older methods measured specific anatomical landmarks or geometric ratios by hand-crafted rules. A DCNN embedding is learned end-to-end from training data, encoding whatever features the network finds most discriminative for identity, which is not directly human-interpretable but tends to be far more accurate.
Why does an embedding similarity score require calibration before use as evidence?+
The raw similarity score from one system is not directly comparable to another system's score or to a probability of a true match, so operational deployments calibrate thresholds against validated benchmarks like NIST FRVT results for the specific system and population in use.
Related terms
- AFRS (India)
- Automated Facial Recognition System, the face recognition module within India's CCTNS, operated by the NCRB and used for searching against mug shots,...
- ENFSI BPM
- The ENFSI Best Practice Manual for Forensic Comparison of Speech (2015, rev. 2022); the operational standard for speaker comparison in European forensic...
- False Positive Identification Rate (FPIR)
- In 1:N identification, the rate at which a probe image of an individual not in the gallery generates a top-rank match. Critical...
- IPD Pixels
- Inter-pupillary distance measured in pixels, the primary resolution metric for face recognition; NIST FRVT finds reliable recognition requires approximately 50+ IPD pixels;...
- Morphological Analysis
- The systematic comparison of qualitative facial feature shapes from a standardised feature list; the primary method in ENFSI BPM facial image comparison,...
- NGI Face Component
- The face recognition module within the FBI's Next Generation Identification system, which searches a gallery of over 600 million images, including driver's...
- NIST FRVT
- The National Institute of Standards and Technology Face Recognition Vendor Testing programme, the primary independent benchmark for facial recognition algorithms, publishing ongoing...
- Photo-Anthropometry
- The measurement of anatomical facial landmark ratios (scale-invariant) from face images; sensitive to head pose differences and requires geometric correction for images...
- Verbal Likelihood Ratio (VLR)
- The scale used by ENFSI and the UK Forensic Science Regulator for forensic comparison conclusions, running from 'No support for the proposition...