Image Augmentation Techniques for Cascade Model Training

Abstract

Cascade classifiers are widely used in embedded systems for tasks of face detection or autonomous driving. To apply cascade classifiers for such advanced tasks, training has to be performed on a large set of samples. Image augmentation methods are introduced to artificially extend training set by creating many altered versions of the same image. This provides more samples to train on, but can also help to expose the classifier to a wider variety of colors and lighting to make a classifier more robust. Due to large number of augmentation methods, selection of appropriate methods depends of data set and classifiers purpose. In this paper the influence of different augmentation methods on traffic light classifier training is investigated. Testing is performed using basic augmentation methods which affect sample's color and luminance characteristics, such as brightness, contrast modifications, blurring and noise addition. Furthermore, horizontal flip as well as color jittering based on principal component analysis were selected as good method candidates while these methods do not significantly influence sample's properties. Each method is applied to a set of training data and the impact it had on the training process is analyzed. Results are presented in the form of rate change of true positives (TP), as well as the rate change of false positives (FP), obtained when comparing different augmentation methods with a non-augmented baseline model. Most of the tested methods reduced the FP count, however, they also resulted in a reduced TP count. The overall results indicate that the selection of augmentation methods heavily depends on the quality of the training data.