Machine and Deep Learning Using MATLAB

8 Regression ClassificationObject Detection

A convolutional neural network (CNN) is often used for image classification. Generally speaking, classification refers to the task of using data and known responses to build a predictive model that predicts a discrete response for new data. For image classification, the input data is the image and the known response is the label of the image subject. On the other hand, regression is another task that can be accomplished with deep learning. Regression refers to assigning continuous response values to data, instead of discrete classes. One example of image regression is correcting rotated images. The input data is a rotated image, and the known response is the angle of rotation.

Preparing Data for Regression

To perform transfer learning for regression, we can use the trainNetwork function with the data, architecture, and training options as input.

NOTE #1

We cannot use an imageDatastore to store training data when performing regression because the labels are not categorical. One convenient alternative is to use the table datatype.The first variable of the table should contain the file directories and names of each image. The second variable should contain the responses. In this case, hand-written synthetic digit images will have one class or output, that is, the rotated angle.

Modification of CNN Architecture from Classification to Regression

Let us load AlexNet and save the network layers in a variable named layers.

>>net = ...