OpenCV Mini Project

This project was made using OpenCV and TKinter (UI library)

Demo

NOTE: Haar Cascade's webcam has an image overlay for experimenting with, please see an indepth explanation below

The goal of this project was to experiment with OpenCV's Thresholding (specifically Gaussian), GrabCut Foreground Extraction, and Haar Cascade. As well as wrapping it up with UI

Project In Detail

• Threshold (Gaussian): The inspiration of Gaussian Threshold operation was by the Firefighter AR : https://www.youtube.com/watch?v=PTfFtsUo7qY I wanted to see for myself if this particular Threshold would have benefitted Firefighters or any related activities

• GrabCut Foreground Extraction: This algorithm, provided by OpenCV made me think about applying image segmentation on an image or a live video feed. Overall this was to help me better understand how it works

• Haar Cascade: Using the exisiting Face and Eye cascades provided in https://github.com/opencv/opencv/tree/master/data/haarcascades, helped me build on top of my previous idea of what I've realized is called "Virtual Fitting". However based on OpenCV's latest version, it has taken out commands to train Haar Cascades.

Experimenting with GrabCut and Haar Cascade

I've used GrabCut to cut around the hat, later manually resized it smaller in order to impose it on the desired image,

When using GrabCut, the final result shows a black background, I need it to be transparent:

b, g, r = cv2.split(img_test_2)
rgba = [b,g,r, alpha]
dst = cv2.merge(rgba,4)

At first, when superimposing the hat on the face, it had excess noise when applying the following code:

ret, mask = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY)

Further research in the documentation shows that an example of applying the gaus filter and otsu to remove excess noise : https://docs.opencv.org/3.4/d7/d4d/tutorial_py_thresholding.html

blur = cv2.GaussianBlur(img2grey, (5,5), 0)
ret, mask = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)

After the image has been finalized, I combine the modifications above to the Haar Cascade

Final Thoughts

Overall a great learning experience, I later realized the training needed to able to perform similiarly to "Virtual Fitting". As well as avaliable third party cascades that would be able to do so.