I am using a ImageNet-trained network to extract features and classify my own images. My images are quite different (microscopic images) from cats and dogs but the features extracted from the ImageNet gave quite promising results by classification.
It would be very easy for me to generate millions of small microscopic images which contain no label. Would it be somehow possible to train my own convolutional neural network? My target images for classification after transfer learning are labeled. Is this somehow possible? Or can I use pseudo labeling (e.g. classes of mean(histogram) or whatever)?