Artificial Data for Image Classification | by Ayaan Haque | Dec, 2020


A novel method for using GANs and artificial data in classification tasks

A schematic of the EC-GAN, or External Classifier GAN, Method
A comparison of real SVHN images and EC-GAN generated images. Left: Real SVHN images. Right: Fake SVHN Images.
A schematic of the common shared architecture discriminator for both classification and discrimination. Image by Thalles Silva (Source)
An illustration of the proposed EC-GAN model. The algorithm is described further below.
Discriminator Loss Objective
Generator Loss Objective
Classifier Loss Objective
DC-Generator Code
DC-Discriminator Code
ResNet Classifier
Example Step by Step of the Algorithm (explained below)
EC-GAN is compared to the shared architecture method on SVHN at different dataset sizes. The left value is the accuracy of a standard classifier (same architecture as GAN counterpart), followed by the accuracy of the GAN classification algorithm.

Read More …


Write a comment