https://github.com/ritu-thombre99/emotion_recognition

Multilayer-Perceptron and CNN for emotion recognition through facial expression along with fuzzy logic.
https://github.com/ritu-thombre99/emotion_recognition

cnn fuzzy-logic mlp mlp-classifier mlp-networks nft

Last synced: 6 days ago
JSON representation

Multilayer-Perceptron and CNN for emotion recognition through facial expression along with fuzzy logic.

• Host: GitHub
• URL: https://github.com/ritu-thombre99/emotion_recognition
• Owner: ritu-thombre99
• Created: 2020-01-07T11:19:49.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2020-01-07T11:24:50.000Z (almost 6 years ago)
• Last Synced: 2025-10-09T01:09:39.577Z (6 days ago)
• Topics: cnn, fuzzy-logic, mlp, mlp-classifier, mlp-networks, nft
• Language: Jupyter Notebook
• Homepage:
• Size: 140 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: readme.txt

Awesome Lists containing this project

README

          
Topic: Emotion recognition through facial expressions

Architecture: Convolutional Neural Network was used

	Input layer with 2304 neurons

	8 hidden layers:

		1) CNN layer with 64 neurons and filter size of 3*3

		2) CNN layer with 64 neurons

		3) CNN layer with 64 neurons

		

		/* Maxpooling after this to reduce 48*48 image to 24*24 */

		

		4) CNN layer with 32 neurons and filter size of 3*3

		5) CNN layer with 32 neurons

		6) CNN layer with 32 neurons

		

		/* Maxpooling after this to reduce 24*24 image to 12*12 */

		/* model.flatten() used to flatten and connect CNN layer to Dense layer

		

		7) Dense layer with 128 neurons

		8) Dense layer with 64 neurons

	Output layer with 7 neurons

	Activation function : Relu

	Activation function for Output layer : Softmax

	

Preprocessing:

	Input was taken from a .csv file and converted to a dataframe with Pandas.

	Pixels were in the form of a string, which were then converted to 2-D INT array of 48*48 using numpy

	

Database:

	Database was a .csv file with : 1) Labels(0-6 representing 7 emotions)

					2) 48*48 pixels in the form of string

					3) Whether data is used for training or testing

					

Experimental analysis:

	a.Number of classes and samples:

				Total number of data : 35,887

				/* Training-testing split of 80-20 */

				Number of training inputs : 28709

				Number of testing inputs : 7178

				Samples : 1) Training

					  2) Testing

					  3) Validation

	

	b.Accuracy : Accuracy obtained was between 50-60%

	

	c.Number of iterations : 30 epochs

	

	d.Improvements: 1) Increasing the number of epochs

	

Furthur notes : Output of CNN was fuzzified i.e POST PROCESING FUZZIFICATION.

		Based on our observations, we defined some FUZZY RULES to furthur improve the accuracy.