Case Study: Tencent Optimizes an Illegal Image Filtering System
- File: tencent-case-study-ipp-illegal-images.pdf
- Size:1016.16 KB
Tencent, China's largest and most used Internet service portal, was able to double the speed of its illegal image filter system using the SIMD instruction set and Intel® Integrated Performance Primitives (Intel® IPP).
For a large Internet service provider like Tencent, being able to detect illegal images is key. Every year, the volume of newly generated images reach about 100 petabytes―even after image compression. Some users may try to upload illegal images (e.g., porn). They certainly don’t tell the system that the image is illegal, so the system runs a check on each image to try to block them. This is a huge computing workload, with billions of images uploaded each day.
Intel helped Tencent optimize the top three hotspot functions of its illegal image filter system. By manually implementing the fingerprint generation with SIMD instructions, and replacing the filter2D function with a call into the Intel IPP library, Tencent was able to speed up both hotspots by more than 10x. As a result, the entire illegal image filter system has more than doubled its speed. This will help Tencent double the capability of its systems. The latency reduction also improves the user experience for customers sharing and sending images.