How to calculate the total number of inputs in CNN?

Assuming you are classifying images (and not pixels or image segments), your samples are the images and your features are the individual pixels. So the Fashion MNIST dataset has $60,000$ samples and each sample has $28 \times 28 \times 1 = 784$ features. It's the number of training samples compared to the number of parameters (which is partly determined by the number of features) that is relevant when considering the potential of a model to overfit.

It is quite common for DL models to have more parameters than training samples and DL models do often overfit. However, there are at least a couple of reasons why this is not necessarily a problem.

1. There are several ways of controlling overfitting in DL models, such as L1/L2 regularisation of the weights and dropout. Using these mean you can have more parameters than training samples without overfitting.

2. There's a phenomenon that has been observed in DL models called "deep double descent" (see Nakkiran et al.'s Deep Double Descent: Where Bigger Models and More Data Hurt, also their Open AI blog post). When plotting test loss against the model size, we see at first the usual descent in the loss curve due to reducing bias as the number of parameters increases, followed by an increase in test loss as the models start to overfit. So far this is as expected. But as the model size is increased even further, the test loss starts decreasing again.

It's probably also worth noting that overfitting in itself is not a problem. It's common for a model to perform slightly better on the training data than the test data and we want the model with the lowest test loss, not the model with the least difference between the training and test loss. Overfitting is only a problem if we haven't allowed for it (e.g. if we make decisions based on the training data results or the model overfits the test data) and so select a sub-optimal model or incorrectly assess the model's ability to generalise.