Aggregation Of Lark Vectors For Facial Image Classification

Face recognition is prevailing to be a key aspect wherever there is a need for interaction between humans and machines. This can be achieved by containing a set of sketches for all the possible individuals and then cross-validating at necessary circumstances. We propose a mechanism to fulfil this task which is centred on locally adaptive regression kernels. A comparative study has been presented at encoding stages as well as at the classification stages of the pipeline. The results are cautiously examined and analyzed to deduce the best mechanism out of the proposed methodologies. All the ideologies have been tested for multiple iterations on benchmark datasets like ORL, grimace and faces 95. The vectorized descriptors have been subjected to encoding using slightly refined methods of feature aggregation and clustering to assist classifiers in imputing the test subjects to their respective classes. The encoded vectors are classified using Gaussian Naive Bayes, Stochastic Gradient Descent classifier, linear discriminant analysis and K Nearest Neighbour to accomplish face recognition. An inference on sparse nature of locally adaptive regression kernels was made from the experimentation. A rigorous study regarding the discrepancies of the performance of LARK descriptors is reported.