Face liveness detection using background/foreground motion analysis
Inventors
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12266215-B2
Publication Date
2025-04-01
Expiration Date
Abstract
Face recognition systems are vulnerable to the presentation of spoofed faces, which may be presented to face recognition systems, for example, by an unauthorized user seeking to gain access to a protected resource. A face liveness detection method that addresses this vulnerability utilizes motion analysis to compare the relative movement among three regions of interest in a facial image and based upon that comparison to make a face liveness determination.
Core Innovation
A method of determining face liveness receives, at a processor of a face recognition system and from an image capture device of the face recognition system, a time-stamped frame sequence. The method generates time-constrained sequential pair-frames by constraining time lapse between frames in the time-stamped frame sequence, identifying corresponding pixels for each pair of time-constrained sequential frames, and segmenting one of each pair of time constrained sequential frames into regions of interest. A motion feature is calculated for each region of interest of each pair of time constrained sequential frames.
The method generates a pair-decision for each pair-frames based on a comparison of the calculated motion features for each region of interest of the pair of time constrained sequential frames. A dynamic decision fusion scheme is applied to make a final liveness determination on all pair-decisions from qualified frames in the time-stamped frame sequence. The dynamic decision fusion scheme is conducted on all pair-decisions to identify changes and patterns, and all pair-decision results are treated together as a time series of decisions.
The approach includes matching face region with pre-enrolled face region images for a genuine check if the user is pre-enrolled. Motion features can be based on estimated pixel velocities, and the regions of interest can be segmented into a face region, a face surrounding region, and a background region. Liveness decisions can be derived from a comparison of motion features across regions and can be based on liveness detection rules stored in a memory.
A training method for a classifier receives time-stamped frame sequences having a live subject and time-stamped frame sequences having a spoofed subject, selects time-constrained frame pairs, identifies corresponding pixels, segments one frame into regions of interest, and calculates a motion feature for each region of interest. The method stores calculated motion features from live sequences as positive data and stores calculated motion features from spoofed sequences as negative data, and generates training rules based on the positive and negative data.
Claims Coverage
The partial content includes three independent claims, covering a liveness-determination method, a training method for a classifier, and a corresponding face recognition system. The independent claims share the inventive core of producing pairwise liveness decisions from region-wise motion features computed over time-constrained frame pairs and fusing the pair decisions via dynamic fusion as a time series.
Time-constrained paired motion-feature liveness determination
Receiving, from an image capture device, a time-stamped frame sequence; generating time-constrained sequential pair-frames; identifying corresponding pixels for each pair; segmenting each pair into regions of interest; calculating a motion feature for each region of interest; generating a pair-decision for each pair based on comparison of the calculated motion features; and applying a dynamic decision fusion scheme on all pair-decisions treated together as a time series to produce a final liveness determination.
Classifier training using positive and negative time-constrained region-wise motion features
Receiving a first plurality of time-stamped frame sequences having a live subject and a second plurality having a spoofed subject; for each sequence selecting a pair of frames that satisfied the time constrain for the time lapse; identifying corresponding pixels; segmenting into regions of interest; calculating a motion feature for each region of interest; storing motion features from live sequences as positive data and from spoofed sequences as negative data; and generating training rules based on the positive and negative data.
Face recognition system with dynamic-fusion liveness from motion-feature pair decisions
A face recognition system comprising an image capture device, a processor, a face detection unit comprising a memory, a face liveness unit comprising a memory, and a face matching unit; wherein instructions cause the processor to receive a time-stamped frame sequence, select a pair of frames satisfying a time constraint, identify corresponding pixels, segment frames into regions of interest, calculate a motion feature for each region, generate a pair decision for each time-constrained pair based on comparison of calculated motion features, and make a final face-liveness determination by applying dynamic fusion on all pair decisions treated together as a time series to identify changes and patterns.
Across the independent claims, liveness is determined by selecting time-constrained sequential frame pairs from a time-stamped frame sequence, extracting region-wise motion features from corresponding pixels, generating pair decisions from comparisons of these motion features, and applying dynamic decision fusion treating all pair decisions as a time series to identify changes and patterns. A related independent claim trains the classifier by generating training rules from positive (live) and negative (spoof) motion features derived using the same time-constrained, pixel-correspondence, region-wise motion-feature processing.
Stated Advantages
Generates a final liveness determination by applying a dynamic decision fusion scheme on pair-decisions treated together as a time series of decisions.
Documented Applications
Determining face liveness in a face recognition system using a time-stamped frame sequence from an image capture device.
Making a genuine check by matching face region with pre-enrolled face region images if the user is pre-enrolled.
Training a classifier for face recognition using time-stamped frame sequences with live subjects and time-stamped frame sequences with spoofed subjects.
Interested in licensing this patent?