Fashion image generation using Conditional Deep Convolutional Generative Adversarial Network based on text input

EasyChair Preprint no. 6505

Abstract

Generative Adversarial Network (GAN) is increasingly becoming a popular research area for generating highly realistic data. However one of the pitfalls of GAN is that it cannot be controlled what a conventional GAN will generate. Therefore we extended a Deep Convolutional Generative Adversarial Network (DCGAN) and added conditioning to it, to generate targeted images according to the user input. We used the fashion MNIST dataset to train our generative model. The text given by the user is matched with predefined fashion categories using fuzzy string matching and the model will generate absolutely new fashion images in accordance with the matched fashion category. The inception score of the generated images are analyzed to evaluate the performance of the model.

Keyphrases: computer vision, deep convolutional network, DNN, Fashion-MNIST, Generative Adversarial Network, image processing

@Booklet{EasyChair:6505,
  author = {Mosarrat Rumman and Abu Nayeem Tasneem and Israt Jahan Ritun and Annajiat Alim Rasel},
  title = {Fashion image generation using Conditional Deep Convolutional Generative Adversarial Network based on text input},
  howpublished = {EasyChair Preprint no. 6505},

  year = {EasyChair, 2021}}