We have seen how MLPs, which receive a single input vector that is then transformed through one or more intermediate hidden layers, can be used to recognize and classify small images such as letters and digits in OCR. However, one limitation of MLPs is their ability to scale with larger images, taking into account not just individual pixel intensity or RGB values, but the height, width, and depth of the image itself.
Convolutional neural networks (CNNs) assume that the input data is of a grid-like topology, and so they are predominantly used to recognize and classify objects in images since an image can be represented as a grid of pixels.
