Image Classification with Variable Input Resolution in Keras

Eric Muccino
Mar 20, 2020 · 4 min read

Convolutional neural networks (CNN) are a type of neural network designed for image classification. For an introduction to CNNs, check out this post by Matthew Stewart.

Problem

Many CNN architectures require that all input images have the same resolution (height and width). This happens when a convolutional layer is flattened and fully connected to a dense layer. Since the network has to initialize the dense layer weights, the convolutional layer resolution must be known, requiring the input…