The Problem
Facial expressions can be difficult to interpret during virtual communication. This project explores whether machine learning can automatically recognize common emotional states.
MACHINE LEARNING PROJECT
Facial Emotion Detection
A machine-learning system that compares multiple neural network architectures to classify facial expressions as angry, happy, sad, surprised, or neutral.
😊
FACIAL ANALYSIS
The Data
Facial images are converted into grayscale pixel arrays and categorized into five emotion classes. Invalid samples are removed before creating training and testing datasets.
The Models
The project compares a pixel-based MLP, a facial-landmark neural network, and a convolutional neural network that learns spatial image features.
The Evaluation
Each model is evaluated using test accuracy, validation loss, training curves, confidence scores, and a five-class confusion matrix.
INTERACTIVE WALKTHROUGH
Explore the classification pipeline
Select a model and sample expression, then run the walkthrough to see how an image moves through the system.
1. Select a model
The convolutional neural network examines nearby groups of pixels to learn facial patterns such as mouth shape, eyebrow position, and eye structure.
2. Select a sample expression
3. Or upload an image
Uploaded images are displayed locally and are not sent to a server.
STAGE 1
Facial Image Input
The model begins with a 48 × 48 grayscale facial image. Reducing the image to grayscale allows the model to focus on facial structure rather than color.
😊
→
DEMONSTRATION RESULT
86%HAPPY
HAPPY86%
NEUTRAL7%
SURPRISE4%
SAD2%
ANGRY1%
These scores demonstrate how model output is visualized. They are not generated by live inference.
MODEL COMPARISON
Three approaches to the same problem
Pixel MLP
Standardized image pixels are flattened into a feature vector and passed through fully connected neural-network layers.
Pixels → Dense Layers → Softmax
Facial Landmarks
Sixty-eight facial coordinates represent important locations around the eyes, nose, jaw, eyebrows, and mouth.
Coordinates → Dense Layers → Softmax
Convolutional Network
Convolution, normalization, pooling, and dropout layers learn progressively more complex visual patterns.
Image → Convolution → Classification