IMPROVING CNN FEATURES FOR FACIAL EXPRESSION RECOGNITION

Karadeniz, Ahmet Serdar (Poznan University of Technology Faculty of Engineering Management)
Karadeniz, Mehmet Fatih
Weber, Gerhard Wilhelm
Husein, Ismail

Zero : Jurnal Sains, Matematika, dan Terapan
Vol 3, No 1 (2019): Zero: Jurnal Sains Matematika dan Terapan

Abstract Facial expression recognition is one of the challenging tasks in computervision. In this paper, we analyzed and improved the performances bothhandcrafted features and deep features extracted by Convolutional NeuralNetwork (CNN). Eigenfaces, HOG, Dense-SIFT were used as handcrafted features.Additionally, we developed features based on the distances between faciallandmarks and SIFT descriptors around the centroids of the facial landmarks,leading to a better performance than Dense-SIFT. We achieved 68.34 % accuracywith a CNN model trained from scratch. By combining CNN features withhandcrafted features, we achieved 69.54 % test accuracy.Key Word: Neural network, facial expression recognition, handcrafted features