Improvement of Facial Expression Recognition Based on Filtering and Certainty Check

The current methods of facial expression recognition have provided various reliable results. However, the accuracy of those methods can be improved by utilizing certain approaches. To improve the accuracy of the detection, this paper proposed two factors (image preprocess and certainty check) that can be used to increase the accuracy of facial expression recognition. To study the efficiency of the two factors, the project firstly built a facial expression recognition script on MATLAB. The code starts with image resizing and duplicating, which resized and duplicated the training and testing images obtained from the FER2013 dataset. The resulting images will used as the input of the pre-trained neural network, AlexNet, to study with specified training options. In the end, the resulting MAT file can be used with Viola-Jones facial detection to classify the emotion shown in the user input figure. To improve the accuracy of the built system, this project implemented the image preprocessing with filters like Difference of Gaussian, median filter, averaging filter, and Gaussian filter, and by comparing the accuracy of the detections with different preprocessing filters, the Gaussian filter has been selected due to its best performance among all the filters. The second approach to higher accuracy is the certainty check before classification. A certainty check would classify the facial expression as “complexed” unless the detection has a possibility of over 85% percent. By combining the two methods stated above, the project has increased the accuracy of detection for happiness by over 40%, and the accuracy for surprise and anger has improved over 10%. The experimental result demonstrats that the project increases its ability in happy emotion detection a lot, and it is beneficial to anger and surprise detection.