Generative Adversarial Network-Based Visual-Aware Interactive Fashion Design Framework

Abstract

Fashion image generation is the task of generating realistic fashion images from a real dataset distribution. Due to the subjectivity of design, fashion industries have always been striving to meet customers’ needs. Although image generation techniques have become more advanced through time, the results are prone to visual-inconsistencies, enormous artifacts, uncontrolled generation, and poor quality at large. This study aims to scale up the image generation process by integrating in the GAN for multiple fashion attributes such as color, shape, and texture to existing architectures. To accomplish this, 12000 Ethiopian fashion images were collected from different sources. As deep learning is a data-intensive approach, image augmentation was used to enlarge the dataset to 90,000. Standard preprocessing was applied to normalize inputs to a common scale, compute average color, create a binary segmented mask, and label the dataset. The conditional inputs added in the proposed architecture are an average color, a segmented binary mask, and a 512 texture dimension organized as Tensorflow records fed to the existing progressive growing GAN generators. The discriminator was assigned to estimate the average color and classify the generated images. Besides, two experiments were done using the same dataset and training configuration namely StyleGAN and it’s conditional version. Improved results were obtained from the three experiments with the evaluation metrics Frechet inception distance, and perceptual path length 41 and 1500, respectively. Moreover, the human evaluation user study was perfomed to assess the user capability on identifying real and generated images and to examine the closeness of these images by using Google forms in both paired and unpaired evaluation settings with the confusion rates of 46% and 47.6% respectively. In general, the performance evalution of implemented conditional progressive growing generative adversarial network along with multiple conditional inputs shows an improved results were achived. As a result, fashion design could be generated using such a GAN methods.