在 go 上重新训练 Tensorflow Model

Question

I m trying to create an application that captures the feed of one camera, detects the faces in the feed, then takes pictures of them and adds them to the image database. Simultaneously another camera feed will be captured and another neural network will compare the faces in the second camera feed with the face images in the database and then will display the name of the person. Ideally, the new face images should be loaded into the neural network model without it completely retraining.

Right now I'm trying to achieve that with TensorFlow and OpenCV.

Would a dynamic neural network be possible with TensorFlow?

Answer 1

As for step two(getting the name of the person), I don't think you would need any retraining to achieve this.

You could use Convolutional LSTM or a similar nn. input shape could be (None,image_dimension_x,y,3) (3 is the color channel, for RGB)

where None would be the current total number of images in the database. It passes all the images in the database into the nn and returns a number as an index.

Or alternatively, you could use a normal convolution (without the None)and make it output the confidence it has for each image in the database to be the person on camera right now. Then choose the person with the highest confidence.

I would say the second one is easier and probably better, that's my suggestion anyway.

Hope it helps:)