A) PIQE/NIQE quality metrics may answer (1)

B) Within the retrieved frames in "A" you may want to use "the Viola–Jones" algorithm to detect faces or you can train a model.

C) (3) more recognizable and clearer are two different objectives. Clear may be attributed to quality, thus you can recalculate PIQE/NIQE from the face region blobs.

Abbas Cheddad that was truly helpful and I appreciate your help.

Do you probably know some method that could detect all the faces in a frame and also give some score to each face that indicates the confidence for correct detection.

have a look at the image if my explanation sounds vague.

Thanks

I know what you meant, and probably the YOLO deep learning method with its different flavors may do the job, there are other attempts to enhance it or to speed-it up (like the algorithm "faced"). Having access to GPU processing would help.

https://github.com/iitzco/faced?source=post_page-----1488681c1602----------------------

If the environment in your videos is not cluttered and you want a simple ML algorithm then you can see the following (uses handcrafted features -HOG- along with SVM)

https://se.mathworks.com/matlabcentral/fileexchange/53849-code-for-face-recognition-with-matlab-webinar

R-CNN could also be an option. Here is the Mathworks' blog on detection of Stop signs that displays the detection confidence level (you can tailor it to your face detection task, you need to get samples of faces for training):

https://se.mathworks.com/help/vision/examples/object-detection-using-deep-learning.html