Python Deep Learning - Second Edition

Data pre-processing

Until now, we've fed the network with unmodified inputs. In the case of images, these are pixel intensities in the range [0:255]. But that's not optimal. Imagine that we have an RGB image, where the intensities in one of the color channels is very high compared to the other two. When we feed the image to the network, the values of this channel will become dominant, diminishing the others. This could skew the results, because in reality every channel has equal importance. To solve this, we need to prepare (or normalize) the data, before we feed it to the network. In practice, we'll use two types of normalization:

• Feature scaling: where . This operation scales all inputs in the [0, 1] range. For example, a pixel with intensity ...