I am trying to create a convolutional neural network for multilabel classification of greyscale images showing the outline of people. I do not have much available data and overfitting is a problem. To counteract this I would like to use transfer learning from a standard conv network, but all the ones I have seen are trained on 3 channel (color) images.
My idea is to 1) transfer the first 8 layer's weights of VVG16 and replace the first layer with a single channel convolutional layer. I'd then 2) freeze the other layers and train the last eight layers. I would then unfreeze all the layers and train again for fine tuning.
Can anyone more experienced tell me if my idea is on the right track or whether it would just be a waste of time? Thanks.