[Lesson 4] Going Further with CNNs (1)

- In the previous lesson, we built and trained a CNN to classify small grayscale images of articles of clothing from the fashon MNIST dataset.

 

- One of the great advantage of CNN's is that they can also work with color images.

 

- Later in the lesson, I'll build and train a convolutional neural network, that can classify color images of cats and dogs.

 

- Along the way, I'll also learn different techniques that can be used to manage a common problem with neural networks called overfitting.

 

- In real applications, we usually have to deal with high resolution color image of different sizes.

 

- In order to decrease the training time, we will only use a small subset of images to train our CNNs.

Data set

 

- When working with this dataset, we will face two main challenges.

1. The first chllenge will be working with images of different sizes.
2. The second chllenge will be working with color images.

- Since all of the images from the Fashion MNIST dataset have the same size, they all get flattened to 1D arrarys of the same size.

 

- Since neural networks need a fixed size input, just flattening the images won't work. Because flattening the images will give rise to one-dimensional arrays of different sizes.

 

- When doing image classification, we always solve this problem by resizing all the images to the same size.

 

- By resizing all images to the same size, this will guarantee that when we flatten the images, they'll result in 1D arrays of the same size.

 

Resizing

 

- Computers interpret color images as three-dimensional arrays. The width and height will be determined by the height and width of the image, and the depth will by determined by the number of color channels.

 

Color Image to 3D array 

 

- Most color images can be represented by three color channels namely red, green, blue. In RGB images, each color channel is represented by its own two-dimensional array.

 

RGB Image

 

- Now, Since our input image is going to be three-dimensional, we have to modify our code accordingly.

ex) input_shape=(150,150,3) <- the third number in the input shape parameter refers to the number of color channels. first two numbers refer to height and width of the input image.