engleski

Efficient facial expression recognition using decision trees and neural networks

This thesis investigates a facial expression recognition system that estimates the emotional state of subjects from facial images. Such systems demand accurate and fast algorithms that can run in real-time on platforms with limited computational resources. The proposed algorithms bridge the gap between precise but slow methods and fast but less precise methods, combining decision trees and neural networks. The gentle boost decision trees are trained to extract highly discriminative feature vectors for each facial expression around distinct facial landmark points. These sparse binary features are concatenated to jointly optimize facial expression predictions with a shallow neural network architecture. The joint optimization improves the recognition rates of difficult expressions such as fear and sadness. Since the algorithm depends on accurate landmark locations, a novel face alignment method is introduced using gradient boost decision trees and neural networks organized in a cascaded regression framework. The cascade is initialized by a lightweight convolutional neural network to increase robustness while preserving high efficiency. The thesis begins with an introduction to the problem and the motivation for solving it, followed by an explanation of the theoretical background and a systematic overview of related, previous work. Next, novel algorithms for face alignment and facial expression recognition are described and evaluated on relevant public data sets. The results demonstrate high efficiency and competitive accuracy compared to the state-of-the-art methods suitable for power-efficient applications. The final chapter provides concluding remarks of the thesis.

decision trees, neural networks, convolutional neural networks, facial expression recognition, face alignment

nije evidentirano