AI-Powered Image Moderation at Meituan: Watermark Detection, Celebrity Face Recognition, and Content Filtering

Meituan leverages AI across its services, building the world’s largest real‑time multi‑point delivery scheduling system and deploying AI‑driven image moderation for millions of daily images.

Background – With millions of images uploaded daily, manual review is infeasible. An intelligent image audit pipeline classifies images as negative (violating) or positive (acceptable) and routes uncertain cases to human reviewers.

Watermark detection – Using an SSD‑based detector built on ResNet, Meituan trained on 1.5 万 water‑mark images (25 classes). Experiments on 3 197 online images show higher recall and precision than traditional sliding‑window methods, especially for unseen watermarks.

Celebrity face recognition – A Faster R‑CNN pipeline with multi‑scale training, context fusion, and Center‑Loss‑augmented Softmax is employed. Ten region‑specific models are ensembled, achieving competitive accuracy on LFW with limited data.

Pornographic image detection – A binary classifier based on deep CNNs (e.g., NSFW‑style) is refined into four categories (porn, erotic, normal person, other). Incremental learning and human‑in‑the‑loop review raise recall while keeping false‑positive rates low.

Scene classification – Transfer learning fine‑tunes a deep convolutional network on Meituan’s vertical categories (food, hotels, etc.). Shallow layers are frozen, deeper layers are retrained, yielding high accuracy on food‑scene and hotel‑room datasets.

References and author biography are provided at the end of the original article.