Facial expression detection using Yolov5

Contents of this time

This time, I would like to write about the facial expression detection program using Yolov5. Please note that it is not possible to post the progress image of the learning data due to the dataset rules.

Model content

Build a model to detect two facial expressions, positive or negative. The model used was: [1]

Based on this data set, we created a positive or negative data set visually.

I knew at the model creation stage before training, but this time I guess that there is a problem that it is easy to approach negative because there is a bias in the category only for female images this time. did.

Learning

part1(epoch50) part2(epoch1000)

The result looks like this.

test

This time, I trained only with images of women, so I tested with 4 images for men and women. part1(epoch50) part2(epoch1000)

Probably the brown one represents negative.

result

This time, there were various biases in the learning model, so the accuracy was not very good, but it seems that facial expression analysis using yolo can be done as an experience, so I would like to make my own model and try again.

Data set used

[1] The Japanese Female Facial Expression (JAFFE) Dataset

Michael J. Lyons, Miyuki Kamachi, Jiro Gyoba.
Coding Facial Expressions with Gabor Wavelets (IVC Special Issue)
arXiv:2009.05938 (2020) https://arxiv.org/pdf/2009.05938.pdf