How Alibaba Detects ‘Disgusting’ Images on Taobao with AI

Abstract

Disgusting images—those that cause nausea—appear in three categories: animal‑related, human‑related, and object‑related. Detecting them among billions of product images faces three technical challenges: few initial samples (cold‑start), severe class imbalance, and highly diverse feature distribution.

Cold‑Start Solution

We expanded the training set by retrieving hundreds of disgusting images from Taobao’s content pool using a small‑sample retrieval platform and applied semi‑supervised learning with a lightweight MobileNet‑V2 backbone. The network output is modeled as Y = f_W(X), where W are parameters and X the input image. A mean‑teacher framework updates the student network while the teacher network tracks an exponential moving average of the student parameters.

Algorithm Iteration Process

Online data exhibits an extreme positive‑negative imbalance (far less than 0.1 % positive). To improve the model we combined active learning, noise‑sample identification, and online hard example mining (OHEM) with cascade training.

Active learning selects samples with confidence between two thresholds (images shown in the figure) and prioritises hard examples for annotation.

The loss‑prediction (LP) module, built on top of the target‑prediction (TP) module, predicts relative loss for each sample pair, guiding the selection of difficult examples.

OHEM + cascade replaces MobileNet‑V2 with DenseNet‑161 as the first stage and feeds its highest‑loss samples to a second ResNet‑50 classifier; inference combines both predictions.

Attention Mechanism

Since disgusting cues are often localized, we embed a CBAM‑style attention block (channel + spatial) into the backbone. Grad‑CAM visualisations confirm that the network focuses on the nauseating regions.

Business Impact

Deployed in Taobao’s “Guess You Like” (首猜) recommendation, the model scans the entire product pool, achieving 95 % precision and 94 % recall. It filtered millions of low‑quality images during the Double‑11 shopping festival, reduced manual review workload by ~70 %, and supported a multi‑task platform for broader image‑quality detection.