Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
RECOGNIZING CHILDREN’S EMOTIONS USING CONVOLUTIONAL NEURAL NETWORKS
Amir Hossein Farzaneh, Xiaojun Qi
DEEP LEARNING
TRAINING IMAGES
TRAINING ARTIFICIAL INTELLIGENCE
NETWORK TOPOLOGY
EXPERIMENTAL RESULTS
SAMPLE APPLICATION
The network is trained using several thousand annotated images of different faces expressing various emotions.
Microsoft’s Facial Expression Recognition (FER+) dataset has 35887 images annotated with 8 emotions.
Challenged with children’s faces detected in a video, the network assigns each child’s emotion in every frame as happy or neutral with high accuracy.
To evaluate the performance of our algorithm, we have manually labelled 1000 images of children’s faces as “happiness” and “neutral”. Our deep learning algorithm classifies images with 88.79% accuracy.
Deep learning algorithms take many forms. Our research group used a convolutional neural network (CNN) to identify eight facial expressions including neutral, happiness, surprise, sadness, anger, disgust, fear, and contempt in people from different demographics and ages. We challenge our network to identify children’s emotions in hope of enhancing Human Computer Interaction (HCI), diagnosing behavioral diseases, and etc.
86.26
85.33
91.67
0 10 20 30 40 50 60 70 80 90 100
Test
Validation
Training
Classification Performance
Accuracy (%)
Evaluation on FER+ dataset:Over multiple iterations, the network discovers patterns in the data that can distinguish emotions. Convolutional layers identify structural features of the images, which are integrated in fully connected layers.
Combining layers of different structure lets the network adapt to recognize images of varying type.
Convolutional layers Fully connected layers
happiness
Input
We adopt from the state-of-the-art CNN architecture VGG13, tweaking the structure of the network to better comply with our application. The output generates probabilities of an arbitrary image belonging to eight distinct classes of emotion. To avoid overfitting, we interleave convolutional layers and fully connected layers with dropout layers.
Confusion matrix
happiness
anger
surprise
Trained CNN
happiness
neutral
This project was supported by USU’s Graduate Research and Creative Opportunities (GRCO) Grant Program 2017-2018