Applying Computer Vision for Content Safety in Live Streaming: Practices and Future Directions

Speaker Zhang Feng, a computer‑vision algorithm engineer at Huya, introduces the use of computer‑vision technology for content safety on the Huya live‑streaming platform.

Project Background: Live‑streaming and short‑video services generate massive amounts of image and video data, inevitably containing prohibited content (political, pornographic, violent, advertising, etc.). Huya employs fine‑grained labeling to quickly recall risky content, achieving second‑level response for billions of items.

Three main moderation approaches are discussed: (1) delayed review – machine flags high‑risk images for human auditors; (2) real‑time interception – machine directly penalizes detected violations; (3) real‑time masking – machine blurs the violating region. Each method’s advantages and drawbacks are compared.

Common Content Risks: Typical risks include political, pornographic, terror‑related, and advertising violations. Examples of regular and difficult pornographic cases are shown, highlighting challenges such as small or hidden subjects.

Image Recognition Algorithm Practice: The "Skyeye" system architecture is presented. Incoming media are processed (screenshot, stream extraction, image preprocessing) and passed through a configurable operator pipeline that outputs tags for downstream actions (human review, real‑time interception, masking). Operators are built and orchestrated to meet specific business needs.

For regular pornographic samples, a multi‑label + multi‑branch classification strategy is used, with manual data augmentation, multi‑tag annotation, and attention‑enhanced backbone networks. For difficult samples, a detection + classification + search pipeline is employed, involving candidate box detection, feature extraction, and similarity search against an indexed library.

Multi‑task learning is applied to handle scarce categories (political, terror) by sharing a backbone across tasks, enabling rapid fine‑tuning of branches and efficient multi‑model deployment.

Operator interfaces are standardized for plug‑and‑play integration. A concrete business example demonstrates how traffic, scene classification, image scaling, and specific violation operators (e.g., game ban, soft‑porn, political‑military) are orchestrated to filter content such as military uniforms, banned games, and soft‑porn.

Future Outlook: The goal is to move from delayed review and interception toward real‑time masking, automatically blurring high‑risk regions without user impact. Examples of real‑time chest‑line masking and text masking are shown.

Q&A Highlights: Questions cover server requirements for porn/terror detection, handling extreme aspect ratios, and text safety methods (OCR + keyword or OCR + NLP).

The talk concludes with a thank‑you and a call to share, like, and follow the DataFunTalk community, which also offers downloadable resources on big data and core AI algorithms.