الفهرس | Only 14 pages are availabe for public view |
Abstract In this thesis, a system for unconstrained face verification based on Hybrid Siamese neural network architecture is proposed. It learns features directly from the face images for face verification and clustering applications. On the unconstrained face image benchmarks, the proposed system provides close to human accuracy on LFW dataset; the system is competitive with stated face verification accuracies as it accomplishes 98.9% under the standard protocol. On Arabian Faces dataset it accomplishes 99.1%. To tackle the cluster quality challenge utilizing the hybrid Siamese neural network architecture, a post-clustering optimization approach is proposed. The proposed post-clustering optimization technique combined with the clustering technique outperforms traditional clustering algorithms as Spectral and K-Means by 0.098 and up to 0.344 as per F1-measure. Experimental results showed that combining the clustering algorithm with proposed post-clustering optimization technique improves the recall and overall F1-measure performance by 0.005 up to 0.219 comparable to the traditional DBSCAN |