Hi,
I have much experiences in AT Retail, customer recognition system based on id recognition,
By using OpenCV's element detector, Dlib's detector, Dib's 68-point land mark feature descriptor, embedding feature extraction using CNN FaceNet, ResNet. and feature classifcaiton using SVM/SVC classification algorithm, it should be processed one by one.
And also for watermarking the alpha channel will be used, and it will be done by cv2 too.
My experience was to identify id card and detect faces using tracking algorithm(OpenCV's cross-correlation tracker, OpenCV's CRAT, MOSSE tracker).
It was deployed on RaspberryPi, Odroid, MiniPC, Jetson devices, Upboards....:
I'm confident I can exceed your expectations.
Thanks,
Evgeny