Understanding convolutional neural networks
In the previous chapters, we have used fully connected Multi-Layer Perceptron (MLP) networks to solve our regression and classification problem. However, as we will see, these networks are not optimal for solving image-related problems.
Images are highly dimensional entities – for example, each pixel in a color image has three features (red, green, and blue values), and a 1,024x1,024 image has more than 1 million pixels (a 1 megapixel image) and, therefore, more than 3 million features (3 * 106). If we connect all these points in the input layer, to a second layer of 100 neurons for a fully connected network, we will require more than 108 parameters, and that would be only for the first layer. Processing images is, therefore, a time-intensive operation.
Furthermore, imagine that we are trying to detect eyes in faces; if a pixel belongs to an eye, the likelihood of nearby pixels belonging to the eye is very high (think of the...
