2020-05-28-face-detection-for-images-using-opencv-part-1.md

layout	cover-img	tags	comments
post	assets/img/blog_images/face-detection-4791810_1920.jpg	deep learning object detection Tensorflow Keras	true

This is part one of this series.

By the end of the series we will build a face mask detector.

For this tutorial we are going to use OpenCV and prebuilt model shipped with it

Face Detection is subset of object detection. It is different from object classification.

The major difference being

Object classification - Given an image, classify this image to some class like Apple, bus, forest etc.
Object detection - Given an image, find out if there exist a patch(or coordinates) where the class exists? eg - Given an image predict whether classes(like oranges, truck, lion) exist in image or not.
source

Since this is an object detection problem,here we need to find the co-ordinates where the face exists in the given image.

To start with head to Github, to download a pre trained caffe model (we would train model from scratch in future parts) and prototxt file (Clone the whole repository to get the complete code)

Make sure you run this command from home dir

pip install -r requirements.txt

This will install all required packages into your local system.

A better practice would be to always have separate environments for different usages (Can be accomplished by using Anaconda)

Now having the setup ready we are good to dive into the code.

Since we will be using console to run this and initialize various objects we would be ArgumentParser The code for which is below

<script src="https://gist.github.com/vipulrai91/2408462b5bb2ab234d7632879d85b59a.js"> </script>

Now, let's load the model in memory and read the images and convert it to OpenCV readable format

<script src="https://gist.github.com/vipulrai91/c68e20ffe1d55d91b0d283f1f8524f16.js"> </script>

Our initial Image looks like this.

Credit: Image by 1138694 from Pixabay

<script src="https://gist.github.com/vipulrai91/2afbfd6ce818a2ee001943ba39bbd14a.js"> </script>

Now we iterate over the detections to extract the confidence (how much percentage the model is confident about predictions) , filtering out weaker confidence , by default it is 0.5 (50%) but this is adjustable.

Finally the output would be the same image with faces marked in boundaries.

This was out of box implementation using pre-trained model provided by OpenCV

In the future posts we would be training model from scratch.

Next up : Using the same model to detect live faces using camera feed.

References :
stackoverflow.com
pyimagesearch.com
Cover Image by Gerd Altmann from Pixabay