Well I think that it depends on the case.

- Are you using each image as one sample?

- How much variability is present in your image domain, and how separable is your solution space?

- How complex is your neural network?

This las point is actually, what most researchers use to estimate their training data size. In general, the complexity of a neural network can be expressed through the number of weights the network needs to adjust. In the case of deep neural networks W can vary from hundreds to millions.

If we call this number W, a good practice is to have ideally at least W^2 data samples. Though it depends on the specifics.

Alvaro is right. It depends on how complex your network is and how difficult your learning problem is.

Having W^2 data samples would be nice, however such a huge dataset is usually merely a wish. Fortunately, significantly less data samples are often sufficient.

In practice, researchers use a validation dataset, which is not used for training. When the classification accuracy of the validation dataset (classification accuracy of unseen data to the network) is comparable to the training performance, then researches hope that the training dataset was sufficient. If the validation accuracy is significantly worse than the training accuracy, the dataset was insufficient.

I agree with Rodriguez. This value varies according to the network complexity. In general, increasing the network's params increases the number of data needed. However, if your data are discriminative as possible, this value could be enough.

Thanks to all of you. Let me study the model as told by you. I will be back to you for further clarification.