LOWERING COMPUTATIONAL BURDEN OF SPARSE REPRESENTATION BASED CLASSIFICATION METHOD FOR FACE RECOGNITION

• Main Author: Indriani L., Susmini
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84505
• Institution: Institut Teknologi Bandung

Face recognition using sparse representation or known as Sparse Representation Classification (SRC) has been a topic that has been continuously developed by many researchers in recent years. This method has been able to address the challenges of face recognition problems such as occlusion or obstructed parts of the face up to 40%, and corrupted image up to 70%. For other challenges such as illumination, expression and pose variation, the SRC method overcomes by adding samples to the training image. However, the addition of samples will burden computation, where this method is already burdened with the problem of solving the minimization solution-1. In this research, the SRC method is developed to reduce the computational burden by reducing the size of the training image dictionary using random projections and combined with clustering techniques to reduce computation. In addition, to reduce computation time, the residual calculation is modified based on the coherence between groups. The hope is that the SRC method will be applicable for the purposes of ubiquitous computing with limited capability computers (low computing devices). By using a random projection matrix (?) can maintain the face recognition performance while a significant reduction in dimensionality is applied, thus the computational burden on SRC-based face recognition can be reduced especially for certain suitable applications that require fast response of face recognition and accept the accuracy performance offered. The best ? matrix should be obtained iteratively in a separate attempt before being used in the SRC algorithm. This best ? matrix is dynamic over different groups of samples, or new samples added to the group may change the best random matrix. Reducing the computational burden is also done by reducing the number of projected training samples to only the closest samples, which are clustered based on the spread between classes derived from Fisher’s criterion. After calculating the cluster with the closest distance to the test sample, the SRC algorithm is then used for the classification process. Tests were conducted on AT&T, Yale B, Georgia Tech, and AR datasets. From the simulation using Python programming language, it shows that random projection combined with clustering method based on scatter will reduce computation time and can even improve accuracy.