Loading...
Please wait, while we are loading the content...
Similar Documents
Heterogeneous face recognition
| Content Provider | Indraprastha Institute of Information Technology, Delhi |
|---|---|
| Author | Sharma, Praneet |
| Abstract | In Heterogeneous Face Recogniton, faces belonging to two di erent modalities are matched. In this research, following two heterogeneous face recognition problems have been considered- VIS to NIR face matching and composite sketch to digital face matching. Face recognition in the presence of illumination changes is still very much an open research challenge. One of the most popular techniques of tackling the problem of illumination changes is to capture the faces in near-infrared (NIR) spectrum and perform matching. But since it isn't feasible to change existing face datasets to all NIR, hence a substitute of this problem becomes NIR to VIS face matching which assumes that though enrollment datasets are in VIS, but the query face images are captured in NIR. This is also known as cross-spectral face matching. Existing algorithms proposed for this problem have been analysed on datasets which don't represent true practicality and have inconsistent evaluation protocols. Thus claims about the performance cannot be made con dently. Matching composite sketches with digital faces is also challenging because of discrepancies in the textural information representation due to errors in witness's description. Automation of this problem would turn out to be quite useful to almost all the law agencies. In this research, cross-spectral framework has been proposed for the problem of NIR to VIS face matching. The framework has also been utilised to study the e ectiveness of variants of Histogram of Oriented Gradients (HOG) features for cross-spectral face matching in conjunction with di erent dimensionality reduction techniques, clasis ers, distance metrics, and fusion tech- niques. For evaluation, NIR-VIS 2.0 dataset has been used. It is observed that the framework outperforms both COTS (commercial-o -the-shelf) system and existing state-of-the-art perfor- mance by 15%. Thus HOG variants are able to successfully represent consistent information across NIR and VIS spectra. Also, directed gradients are observed to be more useful for VIS to NIR face matching. Weighted score fusion is also found to be useful. The later part of the research deals with studying the generalisability aspect of proposed frame- work. In this study, the framework is applied on the task of composite sketch to digital face matching, another heterogeneous face matching problem. The evaluation is done using PRIP composite face dataset. It is observed that the same framework outperforms the existing state- of-the-art by 21%. It is also observed that, unlike in VIS to NIR face matching case, both undirected and directed gradients are able to represent information consistent across two modal- ities. Since this research covers two aspects of heterogeneous face matching, hence is named Hetero- geneous Face Recognition. |
| File Format | |
| Language | English |
| Publisher | IIIT Delhi |
| Access Restriction | Authorized |
| Subject Keyword | HOG descriptor Cross-Spectral Face Matching Weighted Score Fusion |
| Content Type | Text |
| Educational Degree | Bachelor of Technology (B.Tech.) |
| Resource Type | Thesis |
| Subject | Data processing & computer science |
