Facial Image Indexing Using Locally Extracted Sparse Vectors

Amidst all biological characteristics in a human, face emerged to be the most common biometric identifier used by humans due to its distinctive and prominent features. Moreover, the collection of facial images is a friendly, direct and a non-intruding activity, which makes it easily acceptable in nature by the users because of its non-infringement identification technology. We intend to take complete advantage of these developments and propose a novel methodology to achieve identification of the subjects by examining the facial images of the prospects. In the proposed methodology, preprocessed images are passed onto logically adaptive regression kernel for coherent facial feature extraction. The patterns recorded from these feature vectors are condensed using LDA to reduce the computational load. These sparse vectors are quantized and aggregated using VLAD with an intention to classify the descriptors in the later stages of the pipeline. Classification is achieved using CAT Boost and multi-layered perceptron to demonstrate the results using a comparative paradigm. The proposed system has been tested on benchmark datasets such as Grimace, Faces95 and Faces96. Evaluation of these datasets has been done considering the precision, recall and F1-score with an intention to perceive the best one among the proposed alternatives.