I'd need (a lot of) help with MATLAB (facial recognition tasks).

Question

Fortx 2020-2-6

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/503978-i-d-need-a-lot-of-help-with-matlab-facial-recognition-tasks

回答： Harsha Priya Daggubati 2020-4-8

Hello everybody. I need a hand (maybe two) for a facial recognition project I'm trying to do for an university exam. I come from psychological studies, and I'm attending to a course of an Informatics Master's degree. You may ask why, but I like to exit from my comfort zone to learn new things, but sometimes (and that's the case) using this approach you can front some difficulties. I'm sorry if I'll use some incorrect terms.

I have some tasks to do:

localize a face in a new image
check if that face is present in the database I used for train the algorithm
show which subject is most similar to that face from a celebrity dataset given

I collected a dataset of 32 subjects each of them have 10 pictures (so 10 faces).

I found the coordinates of faces on the pictures using Viola Jones (Cascade Object Detector) and using the imcrop, so manually. I saved a .csv containing the coordinates of all the faces.

I used those coordinates to crop the images, so I have all the cropped faces.

I extracted HOG features (providing negatives examples) and classified with SVM, but when I try to evaluate it I don't know what to do. I'd count how many true positive bounding boxes the model found, but how can I do it?

I extracted LBP features, classified whit knn, and measured the accuracy. What if I want to concatenate HOG and LBP features to better train the model? The fact that the number of the columns is very different can be a problem?

I tried to train the Viola Jones, but I don't know how to do it.

Talking about the similarity with the given dataset, how to do it? I thinked about a knn to check the closest to the centroid, but..

Last but not least: how do I make possible to check the whole application using a new image?

Thank you even for reading everything. Some help would be really really appreciated.