IMAGE AUGMENTATION WITH ADDITION OF SIMPLE IMAGE TRANSFORMATION TO GANIMATION RESULTS

• Main Author: Kevin, Christian
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51058
• Institution: Institut Teknologi Bandung

To create an accurate face recognition system, the system needs a lot of human face pictures. In Indonesia, data about Indonesian people’s faces can only be obtained from identification cards. Because of these limitations, it is difficult to make an accurate face recognition system for Indonesian people. The solution offered for this problem is to increase the data of human faces (image augmentation) by adding simple image transformation to GANimation results. GAN is an architecture consisted of two neural networks called generator and discriminator. GAN’s main objective is to reproduce new data that resembles its input. GANimation is a novel GAN conditioning scheme based on Action Units to produce a similar-looking human face with different expressions. GANimation was created by Albert Pumarola and his colleagues in 2019. To increase the amount of image produced by GANimation, the results of image from GANimation will be transformed with simple image transformation. There will be three image transformations used, which are intensity transformation that will alter the image’s brightness, perspective transformation to skew the image, and rotation to rotate the image. In the experiment using face recognition by Adam Geitgey, the image produced by GANimation and simple image transformation can be used and identified correctly by the system. The image produced by GANimation and simple image transformation can also be used for training a face recognition model. There are some limitations in simple image transformation for the images to be able to be recognized by the face recognition system. For rotation, the maximum degree of rotation is 30 degrees. For intensity transformation, the brightness can only be altered half or one and a half of original intensity and for perspective transformation, the maximum shift that can be done is 48 pixels. Outside the limit of these parameter, the face recognition system wouldn’t be able to detect the image as human face image.