Practical Machine Learning for Computer Vision

Afterword

In 1966, MIT professor Seymour Papert launched a summer project for his students. The final goal of the project was to name objects in images by matching them with a vocabulary of known objects. He helpfully broke the task down for them into subprojects, and expected the group to be done in a couple of months. It’s safe to say that Dr. Papert underestimated the complexity of the problem a little.

We started this book by looking at naive machine learning approaches like fully connected neural networks that do not take advantage of the special characteristics of images. In Chapter 2, trying the naive approaches allowed us to learn how to read in images, and how to train, evaluate, and predict with machine learning models.

Then, in Chapter 3, we introduced many of the innovative concepts—convolutional filters, max-pooling layers, skip connections, modules, squeeze activation, and so on—that enable modern-day machine learning models to work well at extracting information from images. Implementing these models, practically speaking, involves using either a built-in Keras model or a TensorFlow Hub layer. We also covered transfer learning and fine-tuning in detail.

In Chapter 4, we looked at how to use the computer vision models covered in Chapter 3 to solve two more fundamental problems in computer vision: object detection and image segmentation.

The next few chapters of the book covered, in depth, each of the stages involved in creating production computer vision machine ...