Augmenting Images with a Mid-Processing Unit to Enhance Classification Accuracy

Abstract

Expression recognition is a challenging task. This paper aims to improve upon the accuracy of an existing Machine Learning classification system, with no-retraining of the existing model, by augmenting the images to improve the classification accuracy. A Mid-Processing Unit is used to manipulate data from the first pass of the classifier, this enhances the original image and improves the overall accuracy result. Three, dimensional reduction algorithms are explored as methods to augment the images; Principal Component Analysis, T-distributed Stochastic Neighbour Embedding, and Non-Negative Matrix Factorisation. Facial Landmarks are also explored as an additional data source. Two phased testing was used; 1. to identify which method combination most improved accuracy, and 2. to fine tune the applied weight to the original images. The final results showed that T-distributed Stochastic Neighbour Embedding in combination with a weight set to 0.024, achieved an almost 1% increase in classifier accuracy.