Advertisement Image Recognition System for 58.com: Design, Implementation, and Performance

Background – With the rapid growth of network bandwidth and multimedia, images and videos have become primary content carriers. 58.com, the largest domestic life‑service platform, processes billions of image requests daily, many of which contain illegal or promotional content that must be filtered.

Motivation – Manual review is costly and inefficient; therefore, an automated solution leveraging recent advances in deep learning for image recognition is essential.

System Capabilities – The Information Security team built a suite of image‑recognition abilities, including object detection, OCR, face recognition, feature matching, and specialized detectors for advertising, pornography, political figures, similar images, ID cards, vehicle models, and license plates.

Advertisement Image Detection – Images are classified as advertisement if they contain text, contact information, watermarks, or QR codes. The system distinguishes three advertisement categories: large‑character posters, regular posters, and QR‑code images; all others are treated as normal.

Solution Design

1. QR‑code Recognition – Detects the presence of QR codes using a mature QR‑code library. Steps: input image → resize while preserving aspect ratio → detect QR code → if found, record position and decode content. The module outputs the QR‑code detection result.

2. Image Classification – A convolutional neural network (CNN) classifies images into large‑character poster, poster, and normal categories. The pipeline includes preprocessing, network architecture design, and category fusion.

Preprocessing – For training, images undergo random cropping, occlusion, blur, and color‑brightness adjustments for data augmentation; in production, images are simply resized to a fixed resolution.

Network Architecture – Based on ResNet, the model adds multi‑scale connections to capture varying text sizes and densities. The architecture diagram is shown in the original document.

Category Fusion – The fine‑grained model predicts dozens of sub‑categories (advertisement types, normal, whitelist). Fusion merges these into the three final output categories.

Output Example

{"QRCode":1.0,"BigCharactorPoster":0.0,"Poster":1.0,"QRCodeContent":"xxxxx","ErrorCode":0,"Normal":0.0}

Accuracy – Sampled evaluations in real‑estate and used‑car scenarios show recall accuracies of 95.8% and 96.1% respectively.

Online Service Metrics – Daily call volume averages ~1,700 × 10⁴ with a peak QPS of 7.69. 90% of calls complete within 18.6 ms, and 99% within 380.3 ms.

Case Studies – The service is deployed in real‑estate, used‑car, and directory mini‑programs, effectively filtering non‑property images, contact information, and other promotional content.

Conclusion – The system provides a low‑complexity, easy‑to‑integrate solution for coarse‑grained advertisement image detection based on text features, with the option to supplement OCR‑based detection for images containing few textual cues.