Mastering Computer Vision with TensorFlow 2.x

Convolution

The convolution is the main building block of a CNN. It consists of multiplying a section of the image with a kernel (filter) to produce an output. The concept of convolution was briefly introduced in Chapter 1, Computer Vision and TensorFlow Fundamentals. Please refer to that chapter to understand the fundamental concepts. The convolution operation is performed by sliding the kernel over the input image. At every location, element-wise matrix multiplication is performed, followed by a cumulative sum over the multiplication range.

After every convolution operation, the CNN learns a little more about the image – it starts by learning the edge, then the shapes in the next convolution, followed by the features of the image. During ...