Wish is a shopping service that puts consumers directly in touch with sellers to buy clothing, electronics, and gadgets at a steep discount. Wish receives around 40,000 User Generated Content (UGC) uploads per day, images which are currently being moderated for upload onto the web platform by human workers. The end goal of this project is to speed up the process of moderating UGC by training a machine learning model to automatically rate if an image is acceptable to view by other users. Images may be rejected for a variety of reasons: NSFW content in the image, the product not being represented in the image, the image having poor quality (e.g. blurry, overexposure, lighting), multiple people or minors appearing in the image, etc. To get a better idea of what this means exactly, the image below on the left would be accepted for upload, but the image on the right would not for the reason stated below. The dataset provided to us by Wish already contains 1 million examples each of accepted and rejected JPEG images for model training purposes. Our goal is to build a model that when given an image as an input, can output a score for acceptability based on image quality, product representation, and automatic rejection criteria. This model will eventually be deployed on an Nvidia DGX-1 containing 8 Volta Cards, with the criteria what it should take less than 5 seconds per image for the model to output a score when running on a GPU.
We are working on designing a Convolutional Neural Network (CNN) to do edge detection and image segmentation so we can better detect blurry images in the foreground.