Design for Embedded Image Processing on FPGAs, 2nd Edition

14Machine Learning

Chapters 7–13 have focused primarily on operations from the low and intermediate levels of the image processing pyramid. Images have been processed to identify regions, and from those regions various discriminating features have been extracted. The final stages in any application are usually classification and recognition. These generally involve models that associate a class or object with the underlying features that have been extracted. The outputs of classification models are usually discrete, with a class label assigned for each possibility that the model has been trained for. Other modelling tasks require a continuous numeric output, for example the quantification of a parameter associated with an object, in which case the modelling is called regression. Machine learning models are often used for these tasks.

The generic architecture of using a machine learning model is illustrated in Figure 14.1. There are two phases to the use of such a model. The first phase is training, where the model parameters for the task are determined. Training is based on datasets that consist of many example images or feature vectors. Training can be categorised as supervised or unsupervised learning. With supervised learning, training uses labelled data, where the corresponding classification or regression outputs for each training example are known. Unsupervised learning looks for patterns within the training dataset and uses these to represent or categorise the features. ...