Visual Algorithm Applications in Advertising Scenarios

AI algorithms, especially deep learning, have rapidly advanced and are now widely applied to provide intelligent solutions across domains. At the 2019 Tencent Advertising Algorithm Competition final, senior researcher Shi Ruichao presented "Visual Algorithm Applications in Advertising Scenarios," focusing on how visual algorithms are used within Tencent Advertising.

Ad Creation – Assisted Creative Generation

Ad creation is divided into image and video creation. Two main challenges for image creative generation are edge handling for product cut‑out and the expansion of template libraries. Tencent Advertising adopted a GCN‑based edge processing algorithm with weighted loss to improve edge smoothness and visual quality.

For template generation, the system extracts the dominant color of the product, retrieves and modifies backgrounds, applies a deep aesthetic filtering model, and adjusts color schemes to produce diverse template materials, achieving a 91% usable rate.

Video creative generation addresses the shortage of video ad inventory. By extracting product information, generating storyboards from an industry template library, rendering videos, and automatically creating cover images, the end‑to‑end pipeline dramatically reduces manual effort while delivering effective video ads.

Ad Review – Stronger Quality Control

Traditional manual review cannot keep up with growing ad volume and complex rules. Tencent Advertising introduced an "Intelligent Review" project that uses algorithms to pre‑detect violations before human review, ensuring consistent feedback.

Examples include using an Asoftmax Loss algorithm to detect celebrity face plagiarism and key‑region image matching for game plagiarism detection. To further improve efficiency, the AutoML engine is employed for rule tuning and search strategy optimization.

Ad Playback – User‑Centric Experience Optimization

The playback stage focuses on increasing click‑through and conversion rates while enhancing visual experience. A dual‑tower model captures historical material click features and combines ad and user features to predict CTR, yielding a 2% AUC lift and 3‑4% online performance gain.

Additionally, a pre‑trained CNN extracts convolutional features, reduces them to a 152‑dimensional vector, and fine‑tunes with similar‑ad data to improve recall and avoid repetitive ad material.

Overall, visual algorithms have significantly optimized Tencent Advertising's workflow, opened new research directions, and the team invites more AI talent to explore deeper applications in advertising.