2024 Bilibili Technology Patent Awards – Highlights of Ten Winning Innovations

On April 26, 2024, the 23rd World Intellectual Property Day marked the end of China’s National Intellectual Property Awareness Week. Bilibili organized an internal 2024 Bilibili Technology Patent voting activity and selected ten outstanding technology patents.

Large‑Scale Video Duplicate Detection Algorithm A two‑stage strategy is employed: the first stage performs coarse filtering to reduce the candidate set and computational load, while the second stage searches a loss matrix generated from query‑video and registered‑video fingerprint vectors to find precise segment matches. This approach can complete segment matching within 10 seconds even against a billion‑scale vector retrieval library.

AI Story Generation System The system automatically generates story text from selected images or videos, then produces corresponding voice‑overs, subtitles, transitions, background music, and filters. It incorporates keyword recognition, text risk control, large‑language‑model expansion, text‑to‑speech, sentence matching, and style‑package‑driven timeline construction, significantly lowering the barrier to video creation.

Live Interactive Rhythm‑Game Play The solution streams live video with embedded interactive rhythm‑game information. While the client plays the video, the interactive game is displayed and responds to user actions, achieving synchronized interaction between live broadcast and gameplay to enhance user experience.

Automatic OTT Service Launch By adding vendor‑specific fields to the DLNA device information protocol, the casting function triggers the corresponding OTT service on the DMR device. The OTT service receives a URL, parses it for playback, and then seeks a DMC device to establish a connection, automatically launching the appropriate casting service and improving the casting experience.

Community Content Governance Applied to community moderation, the system extracts topics from comments and bullet‑screens, builds a knowledge‑graph of recurring issue types, and during content inspection references this graph to automatically identify and control problematic posts.

Live Multi‑Audio‑Track Interactive Play This solution addresses multi‑audio‑track switching in live streams by using a mapping table to correct track identifiers, ensuring data consistency and smooth transitions without audio glitches, thereby improving live‑stream quality.

Bilibili Unified Video Playback Page Integration The architecture decomposes complex business logic into cohesive modules aligned with real lifecycle and driving relationships, using dependency injection to reduce code volume and accelerate development.

Dynamic Resolution Video Encoding Designed for video encoding, a multimodal neural network takes both encoding features and extracted frame sequences as inputs. It predicts the optimal encoding resolution and parameters (RF) that achieve the best bitrate for a given quality target measured by VMAF.

Reference Frame Pruning In AV1 encoding, the method evaluates intra‑ and inter‑prediction rate‑distortion costs during pre‑analysis to decide whether a frame prefers forward or backward reference. It then prunes the less‑preferred direction, reducing the number of reference frames traversed in the main encoding loop and improving efficiency.

GPU Isolation and Sharing Through the BGM module, user‑specified memory and compute limits are set via cgroup interfaces. Drivers query these limits through BGM APIs, enforcing memory and compute restrictions and feeding back to the BGM kernel module, thereby reducing performance overhead.

*Note: The ranking does not reflect any order of precedence.