SYNTHETIC POSITIVE IMAGE GENERATION FOR FINE GRAIN IMAGE SIMILARITY BASED APPAREL SEARCH

In apparel search context, process of finding a similar item out of thousands of other items is a cumbersome and computationally heavy process. In order to build a deep learning model that can perform the similarity search, hundreds of training images per Stock Keeping Unit (SKU) are required. Due to shortage of training data, this approach fails to generate a deep learning model that can perform the similarity search in intended manner. The existing approaches may also require domain experts to perform classification of apparels, so as to generate the training data. The method and system disclosed herein provide an approach in which positive images and negative images are generated from each query image, which in turn are used for generating a training data. The training data is then used to generate a deep learning model, which is used to perform the similarity search.