The power of Machine Learning applied to Image Recognition with CNN

Example of Image Recognition

Image Recognition is the process of identifying a feature or a object in an image or a video (that basically is a sequence of image where time is the third dimension). This concept is used in many processes for automation and it can be applied in several fields as process monitoring, surveillance, medical diagnosis and so on.

Image Recognition is made possible with the CNN, Convolutional Neural Network that are becoming also more popular than the ANN, Artificial Neural Network as shown in the graph below (Google trends)

Trend of the last 5 years of ANN vs CNN

But how does CNN works? It takes an image as input, and, after the input processing, the output will be the classification of the image. In order to classify the image, we see the image as an array where every single element is a pixel.

If we consider a coloured image 4x4 pixels, we have to consider 3 array for RGB colours (Red, Green and Blue). Every single value of every array can be from 0 to 255.

Array for coloured image 4x4 pixels (RGB)

It is simpler for a black and white image because in this case we would have a single array with 0 to 255 possible value.

The steps of our classification will be:

1. Convolution
2. Pooling
3. Flattening
4. Full Connection

Step 1: Convolution

Let’s make it simple making an example :)

We have an input image for example an array of 10 x 8 pixels and we have also a Feature Detector that is an array of 3 x 3 pixels. These two arrays multiplication generate a third array that is the Feature Map that will be an array of 8 x 6 pixels (because it contain the 48 possible combination of 3x3 array inside the 10x8 array). This Process is the Convolution that generates convolutional layers.

This make the image “smaller” and easier to processate. With this procedure we loose…