The key idea of keras-vis visualization for Dense layers is to generate an input image that maximizes the final Dense layer output corresponding to the bird class. So in reality what the module does is reverse the problem. Given a specific trained Dense layer with its weight, a new synthetic image is generated which best fits the layer itself.

A similar idea is used for each conv filter. In this case, note that the first ConvNet layer is interpretable by simply visualizing its weights, as it is operating over raw pixels. Subsequent conv filters operate over the outputs of previous conv filters, and therefore visualizing them directly is not necessarily very interpretable. However, if we consider each layer independently we can focus on generating only a synthetic input image which maximizes the filter output.