Dealing with unbalanced data: Generating additional data by jittering the original image
In many deep learning applications, we often come across data sets where one type of data may be seen more than other types. For example, in a traffic sign identification task, there may be more stop signs than speed limit signs. Therefore, in these cases, we need to make sure that the trained model is not biased towards the class that has more data. As an example, consider a data set where there are 5 speed limit signs and 20 stop signs. If the model predicts all signs to be stop signs, its accuracy is 80%. Further, f1-score of such a model is 0.88. Therefore, the model has high tendency to be biased toward the ‘stop’ sign class. In such cases, additional data can be generated to make the size of data sets similar.
One way to collect more data is to take the picture of the same sign from different angles. This can be done easily in openCV by applying affine transformations, such as rotations, translations and shearing. Affine transformations are transformations where the parallel lines before transformation remain parallel after transformation.
Below I present a function that can be used to generate jittered images. The function takes in the original image, range of angle rotation, range of translation and range of shearing and returns a jittered image. As the function chooses true transformation values from a uniform distribution that is specified by these ranges.
Note, the same techniques can be applied using image generator in Keras. However, the transform_image() function provided here will help you play with the parameters and see whats happening under the hood.
After applying the function transform_image() found at . The new generated images are shown in the figure below.
The code to execute the command above can be found here, https://github.com/vxy10/ImageAugmentation