Multidimensional Neuro-fuzzy System and Fuzzy Coding for a Constant Length Facial Landmark Set Formation
Yevgeniy V. Bodyanskiy, Nonna Ye. Kulishova · 2021 11th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS) · 2021
Facial expression recognition is performed using deep and machine learning. Deep networks guarantee recognition accuracy when large, pre-labeled image sets are available for learning. This condition is not always met. Machine learning approaches will be more useful, especially in real-time tasks, under changing shooting conditions. The main problem in this case is formation of a facial landmarks coordinates vector of constant dimension. Input vector fixed size is one of machine learning algorithms main requirements. However, standard computer vision detectors detect a different number of feature points in each video frame. This paper considers problem of constructing a facial landmarks coordinates vector, which will have a constant length. For this, a fuzzy encoder based on flexible two-dimensional membership functions is proposed, as well as a multidimensional neuro-fuzzy system for encoded data classification.