[Lesson 4] Going Further with CNNs (2)

- In a previous lesson, we learned how to perform convolutions on grayscale images. But, how do we perform a convolution on a color image?

 

- To perform a convolution of this kernel with the given grayscale image, we center our kernel over the pixels of the image, and take each corresponding pixel and kernel value, multiply them together, sum the whole thing up.

 

- Remember that we can add zero padding around the entire image in order to calculate the kernel convolution of the image withour losing information.

 

- In color image, Just as we did with grayscale images, we'll start by choosing a filter of a particular size. 

 

3D Filter

 

- It's customary to add a bias value, which usually has a value of one.

 

Convoluted output

 

-  In this particular example, the resulting convoluted output is a two-dimensional array with the same width and height of the RGB image.

 

- A convolution with a single 3D filter produces a single convoluted output.

 

- However, when working with CNNs, it's customary to use more than one 3D filter. If we use more than one filter, then we will get one convoluted output per filter.

 

3D convoluted output

 

- In this example, since we are using three filters, the convoluted output will have a depth of three.

 

- For example, in order to create three filters with a size of three by three like we have in our example, The code is

ex) tf.keras.conv2D(3, (3,3), ...)

 

- Remember that when we train the CNN, the values and the three-dimensional kernels will be updated so as to minimize the loss function.

 

- In essence, max-pooling works the same way as it did for gray-scale images. The difference is that now we have to perform max pooling on each convoluted output.

 

Max Pooling

 

- In this particular case, when we perform max-pooling on the 3D convoluted output, we get a 3D array that is half the width and height of the convoluted output, but has the same depth.

 

Result of Max Pooling

 

<Colab Notebook>

- To access the Colab Notebook, login to your Google account and click on the link below:

Cats and Dogs

Google Colaboratory

 

<소스 코드>

github.com/HoYoungChun/TensorFlow_study/blob/master/05_Dogs%20vs%20Cats%20Image%20Classification%20Without%20Image%20Augmentation.PY

HoYoungChun/TensorFlow_study