Hardware architecture design for real-time SIFT extraction with reduced memory usage

Tsung Han Tsai, Rui Zhi Wang, Nai Chieh Tung

Research output: Contribution to journalArticlepeer-review


Scale-invariant feature transform (SIFT) is considered one of the best algorithms to get feature points in an image. It maintains the accuracy in results even in image scaling, rotation, deformation, and light changes. However, memory requirements are the bottleneck to achieving real-time performance as SIFT has high computation complexity. Therefore, this paper has proposed the improved hardware architecture of the scale-invariant feature transform (SIFT) algorithm. The Gaussian pyramid is constructed using parallel operations instead of the original cascade operation to reduce memory requirements. Coordinate Rotation Digital Computer (CORDIC) has been adapted to perform trigonometric operations to reduce computational complexity. The ASIC design has been implemented using TSMC 90 nm technology. The system can achieve the performance of 35.6 FPS for an image resolution of 1280 × 720 while using only 237.4 Kbit of memory.

Original languageEnglish
Pages (from-to)6297-6317
Number of pages21
JournalMultimedia Tools and Applications
Issue number2
StatePublished - Jan 2024


  • Gaussian pyramid
  • Key-point
  • Low memory VLSI design
  • Object recognition
  • Real-time


Dive into the research topics of 'Hardware architecture design for real-time SIFT extraction with reduced memory usage'. Together they form a unique fingerprint.

Cite this