A short feature vector for image matching: The Log-Polar Magnitude feature descriptor

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Optimal feature detector-descriptor selection is crucial for image matching, often varying with application and image type.
• Existing methods like SIFT and SURF have limitations in terms of feature vector length and computational requirements.

Purpose of the Study:

• To propose a novel feature descriptor, the Log-Polar Magnitude descriptor, that is invariant to rotation, scale, and illumination.
• To achieve comparable or superior performance to existing descriptors like SIFT and SURF with significantly shorter feature vectors.
• To develop a detector-independent descriptor suitable for applications with limited hardware capacity.

Main Methods:

• The Log-Polar Magnitude descriptor is based on the Log-Polar Transform followed by a Fourier Transform.
• Magnitude spectrum components are selected, allowing optimization for specific image patterns.
• The descriptor relies on feature coordinates and sizes, making it detector-independent.

Main Results:

• The proposed descriptor achieves performance comparable to SIFT on the Oxford dataset and outperforms SIFT on microscopy datasets.
• It generates shorter feature vectors (48 or 56 dimensions) compared to SIFT, improving memory usage and matching speed.
• Rotation invariance is achieved through the magnitude spectrum of the Log-Polar Transform, eliminating the need for orientation estimation.

Conclusions:

• The Log-Polar Magnitude descriptor is a highly efficient and effective alternative for image registration tasks, particularly in microscopy.
• Its reduced feature vector size and computational efficiency make it attractive for resource-constrained applications.
• The descriptor's robustness and performance across diverse datasets highlight its versatility in computer vision.