Source-linked AI summary
Face Identification by SIFT-based Complete Graph Topology
Dakshina Ranjan Kisku, Ajita Rattani, Enrico Grosso, Massimo Tistarelli
TL;DR
Face identification must contend with substantial appearance variation, and prior SIFT-based face authentication motivates more robust matching. This paper matches complete graphs built from face-image SIFT features using two constraints, and reports increased performance over previous SIFT-based work on BANCA.
Problem
Face identification is challenging under imaging and appearance variations, while prior SIFT-based face authentication results motivate more robust matching techniques.
Source: §I · p. 1Method
The system extracts SIFT features from face images, represents each face as a complete graph, and applies gallery-image-based or reduced-point-based matching constraints.
Source: §I · p. 1, §III · p. 2Results
On BANCA, the paper reports a remarkable performance increase over previous SIFT-based work and capability to cope with illumination changes and occlusions.
Source: §V · p. 5Takeaways & Limitations
The reported identification accuracy indicates that improved matching techniques can increase system performance within the same feature representation space.
Source: §V · p. 5
Abstract
from arXiv · showhide
This paper presents a new face identification system based on Graph Matching Technique on SIFT features extracted from face images. Although SIFT features have been successfully used for general object detection and recognition, only recently they were applied to face recognition. This paper further investigates the performance of identification techniques based on Graph matching topology drawn on SIFT features which are invariant to rotation, scaling and translation. Face projections on images, represented by a graph, can be matched onto new images by maximizing a similarity function taking into account spatial distortions and the similarities of the local features. Two graph based matching techniques have been investigated to deal with false pair assignment and reducing the number of features to find the optimal feature set between database and query face SIFT features. The experimental results, performed on the BANCA database, demonstrate the effectiveness of the proposed system for automatic face identification.
I. INTRODUCTION
Face identification is difficult under appearance and imaging variations, and prior approaches leave gaps in capturing facial variation or robustly matching SIFT features. The paper applies graph matching to SIFT features using two matching constraints and evaluates the system on BANCA.
I. INTRODUCTION
- Face identification is challenged by illumination, clutter, scale, translation, rotation, pose, expression, and occlusion variations that alter facial appearance.Source: §I · p. 1
- Appearance-based component-analysis methods are described as inefficient at capturing substantial facial variation or new-class samples, while a cited elastic graph-matching system was not tested under different constraints and made identification complicated.A cited probabilistic approach addressed imprecise localization, partial occlusion, and extreme expressions, but illumination remained unsolved.Source: §I · p. 1
- Although SIFT features had been explored for face authentication with promising results, the paper identifies a need for more robust matching techniques to improve overall performance.The passage notes that the illumination problem remained unsolved for the cited probabilistic approach.Source: §I · p. 1
- The proposed system photometrically normalizes face images, extracts rotation-, scale-, and translation-invariant SIFT features, then matches faces using gallery-image-based or reduced-point-based constraints.Experiments use the BANCA database with the MC protocol.Source: §I · p. 1
II. INVARIANT AND ROBUST SIFT FEATURES
SIFT represents local image features designed to remain stable across several image transformations and is extracted through staged detection, localization, orientation assignment, and descriptor calculation. The paper describes gradient-based orientation and descriptor construction, including an eight-plane, 4 x 4 sampling scheme.
II. INVARIANT AND ROBUST SIFT FEATURES
- SIFT is described as invariant to rotation, scaling, translation, partly illumination changes, and projective transform, with stable points detected through staged filtering in scale-space.Affine-invariant features are also described as distinctive and likely to match under image distortions and illumination conditions.Source: §II · p. 1
- Feature extraction selects scale-space peaks from a difference-of-Gaussian function, localizes stable points, assigns orientations, and calculates descriptors for local shape distortions and illumination changes.Source: §II · pp. 1–2
- Each feature receives a consistent orientation from local image properties, and its orientation histogram summarizes gradient directions within a circular window.Nearby data are fitted to refine location, edge response, and peak magnitude.Source: §II · p. 2
- For illumination invariance, gradient magnitude and orientation are smoothed and sampled over a 4 x 4 grid with 8 orientation planes.Source: §II · p. 2
III. REPRESENTATION OF FACES
The paper represents each face as a complete graph over SIFT feature points and investigates two matching constraints. Its correspondence definitions distinguish directional matches from one-to-one matches intended to omit false assignments.
III. REPRESENTATION OF FACES
- Each face is represented by a complete graph drawn on SIFT feature points, and the proposed constraints seek a corresponding subgraph in the probe image given the gallery graph.Source: §III · p. 2
- Directional correspondence between feature points does not imply the reverse match, whereas one-to-one correspondence requires conditional-probability conditions in both directions to omit false matches.Source: §III · p. 2
- A correspondence graph records one-to-one node pairs between the two face graphs.Source: §III · p. 2
- The paper develops and investigates two match constraints based on its graph and correspondence definitions.Source: §III · p. 2
B. Graph Matching Methodologies
The methods represent face images as complete graphs over SIFT features and match gallery features to a probe using constraints intended to reduce false assignments. They compare graph dissimilarity and also weight feature reliability.
B. Graph Matching Methodologies
- The gallery-image constraint assumes corresponding points occur near similar face positions and uses minimum Euclidean distances to resolve repeated assignments.The method computes node dissimilarities between complete graphs and uses Hausdorff distance for the distances.Source: B. Graph Matching Methodologies · pp. 2–3
- Complete graphs over SIFT feature points support matching a gallery face subgraph to a probe face.Each feature point carries location, descriptor, scale, and orientation information.Source: B. Graph Matching Methodologies · pp. 2–3
- Graph dissimilarity combines vertex and edge scores, with edge distances averaged over corresponding edge pairs.The method constructs complete graphs after finding corresponding feature points.Source: B. Graph Matching Methodologies · p. 3
- The reduced-point constraint selects the minimum-distance match for multiple assignments and removes links without a reciprocal assignment.It addresses both repeated matches to one point and one-way assignments.Source: B. Graph Matching Methodologies · pp. 3–4
- Reduced-point matching is described as more efficient because it matches a small number of feature points with few floating features.Source: B. Graph Matching Methodologies · p. 4
- Gaussian empirical-rule weighting assigns feature weights based on relevance, with node and edge means and standard deviations computed for each image pair.The method uses tested multipliers of 0.075, 0.05, or 0.025 for values within one, two, or three standard deviations, respectively.Source: B. Graph Matching Methodologies · p. 4
IV. EXPERIMENTAL RESULTS
The BANCA evaluation reports prior equal rates and weighted error rates for the two graph-matching constraints, with ROC curves for G1 and G2. The authors state that complete-graph context information is advantageous.
IV. EXPERIMENTAL RESULTS
- Testing used BANCA's Matched Controlled protocol, training on session one and testing on sessions two through four across two groups of 26 subjects.Client-specific thresholds were set using G1 scores, then used to compute false acceptance and false rejection rates on G2; WER uses R as the false-acceptance/false-rejection cost ratio.Source: §IV · pp. 4–5
- The evaluation presents prior equal rates for G1 and G2 and weighted error rates for the two graph-matching techniques at three values of R.The corresponding ROC curves are shown for G1 and G2; the supplied passages do not provide table cell values.Source: §IV · p. 5, Table I · p. 5, Table II · p. 5
- The reported results indicate an advantage from incorporating context through complete-graph topology.Prior equal rates and weighted error rates are presented for both matching constraints, with ROC curves for G1 and G2.Source: §IV · p. 5, Table I · p. 5, Table II · p. 5
V. CONCLUSION
The paper proposes two SIFT-based complete-graph matching methods for face identification and reports improved performance over prior SIFT-based work. It also reports capability to cope with illumination changes and occlusions in the database or query image.
V. CONCLUSION
- The proposed face-identification system builds complete graphs from SIFT features and uses two matching constraints to address false pair assignments and select feature matches.The methods were tested on BANCA using the MC protocol.Source: §V · p. 5
- The authors report a remarkable performance increase over previous SIFT-based work and capability to cope with illumination changes and occlusions.These claims concern the BANCA evaluation.Source: §V · p. 5
- The reported EBGM errors are higher, but the EBGM experiments used FERET and Bochum databases, and direct comparison is left for future testing.The authors note that EBGM also reports recognition across different poses.Source: §V · p. 5
- The paper reports that improved matching techniques can increase system performance within the same feature-representation space.Source: §V · p. 5