A Study on Image Retrieval System for Clothing Materials Using Convolutional VAE

Authors

  • Yeonghun Lee
  • Jonghyung Sung
  • Hyunghwa Ko
  • Kyounghak Lee

Abstract

Recent Image Processing with Deep Learning has been permeated in our daily life. In accordance with development of Deep Learning, we attempt to apply it to Clothing Materials Retrieval.
Proposed model consists with YOLO detecting clothing materials and Convolutional VAE characterizing images in database. Since the vectorized image feature is high-dimensional, PCA is applied to reduce features of image and time for retrieval. Furthermore, KNN is utilized to search for k-similar images in reduced vector space. To test this system, we collected dataset by Web Crawling and its result shows 10-near images for each arbitrary test image.
YOLO is efficient model to produce an image with exactly necessary region only. As an Object Detector, it achieves to generate cropped images appropriately from poor dataset. VAE needs huge dataset in training stage and Encoder of VAE amply works as an image discriminator. Even though we reduced dimension of vectorized features from 512 to 73 or lower dimensions by PCA, it does not have terrific loss. It rather brings reduction of an image search response. KNN is suitable distance measure method when sort out similar images in comparatively low dimension. Importance of our study is on implementation of clothing materials retrieval system by combining Object Detection as region of interest, unsupervised Deep Learning, PCA and KNN algorithms. Proposed model favorably retrieves similar clothing materials based on color and texture without any labeling process in real-time.
Our image retrieval system for clothing materials results 0.1 second for 10-similar image retrieval from 170,681 images on 73-dimensional latent space.

Downloads

Published

2020-03-26

Issue

Section

Articles