Source-linked AI summary
Image Matching Using SIFT, SURF, BRIEF and ORB: Performance Comparison for Distorted Images
Ebrahim Karami, Siva Prasad, Mohamed Shehata
TL;DR
Robust image matching must handle transformations and deformations while balancing matching quality and computational cost. The paper compares SIFT, SURF, and ORB by transforming original images and measuring keypoints, matching rates, and execution times. ORB is fastest, while SIFT performs best in most scenarios, with ORB and SURF leading near 90° rotations and ORB and SIFT performing similarly on noisy images.
Problem
The paper addresses the need to determine how SIFT, SURF, and ORB perform under scaling, rotation, noise, fish-eye distortion, and shearing in image matching.
Method
The study manually applies transformations to original images and evaluates keypoint counts, matching rates, and execution times for SIFT, SURF, and ORB.
Results
ORB is fastest, while SIFT performs best in most scenarios; ORB and SURF outperform SIFT near 90° rotations, and ORB and SIFT perform almost similarly on noisy images.
Takeaways & Limitations
Algorithm performance depends on the distortion: SIFT is generally strongest, ORB is fastest, and rotation near 90° favors ORB and SURF.
Abstract
from arXiv · showhide
Fast and robust image matching is a very important task with various applications in computer vision and robotics. In this paper, we compare the performance of three different image matching techniques, i.e., SIFT, SURF, and ORB, against different kinds of transformations and deformations such as scaling, rotation, noise, fish eye distortion, and shearing. For this purpose, we manually apply different types of transformations on original images and compute the matching evaluation parameters such as the number of key points in images, the matching rate, and the execution time required for each algorithm and we will show that which algorithm is the best more robust against each kind of distortion. Index Terms-Image matching, scale invariant feature transform (SIFT), speed up robust feature (SURF), robust independent elementary features (BRIEF), oriented FAST, rotated BRIEF (ORB).
I. INTRODUCTION
The paper motivates robust feature detection and image matching for computer vision and robotics, then compares SIFT, SURF, and ORB across multiple image distortions.
- Feature detection abstracts image information and makes a local decision at each image point about whether a feature exists.
- Ideal features should remain robust under rotation, scale, illumination, noise, and affine transformations while remaining highly distinctive.
- SIFT is effective for object recognition but has high computational complexity, limiting its suitability for real-time applications.
- SURF approximates SIFT and is faster without reducing detected-point quality, while BRIEF lowers complexity with almost similar matching performance.
- The paper compares SIFT, SURF, and ORB under rotation, scaling, shearing, fish-eye distortion, and noise, evaluating key points, matching rate, and execution time.
II. OVERVIEW OF IMAGE MATCHING TECHNIQUES
The paper outlines how SIFT, SURF, and ORB detect and describe local image features, with SURF approximating SIFT and ORB combining FAST with a rotation-steered BRIEF descriptor.
- SIFT: SIFT estimates scale-space extrema with Difference of Gaussian, refines keypoints, assigns orientations, and generates gradient-based descriptors.
- SURF: SURF replaces Gaussian averaging with faster box-filter convolutions using integral images and detects points with a Hessian-based BLOB detector.
- SURF: SURF assigns orientations and constructs descriptors from horizontally and vertically oriented wavelet responses with Gaussian weighting.
- ORB: ORB combines the FAST detector with the BRIEF descriptor, selects top points using a Harris measure, and estimates orientation from intensity-weighted patch centroids.
- ORB: ORB steers BRIEF descriptors using a rotation matrix derived from patch orientation to improve performance under in-plane rotation.
III. SIMULATION RESULTS
The experiments compare SIFT, SURF, and ORB under intensity, rotation, scaling, shearing, fisheye distortion, and noise. Matching performance varies by distortion: SIFT leads for intensity, selected rotations, shearing, and fisheye distortion; ORB leads for scaling and is fastest under varying intensity; SIFT and ORB are strongest under salt-and-pepper noise.
- Intensity variation: SIFT provides the best matching rate for varying-intensity images, while ORB provides the least and requires the least computational time.The varying-intensity comparison uses Tables 1 and Figure 1.
- Rotation: At 45° rotation, SIFT achieves a 65 % matching rate.The 45° rotated-image results are reported in Table 2 and Figure 2.
- Rotation: ORB and SURF provide the best matching rate at rotation angles proportional to 90°, whereas SIFT leads at 45°, 135°, and 225°.Table 3 compares matching rates across rotation angles.
- Scaling: ORB achieves the highest matching rate for an image scaled by 2 times, while SIFT achieves the least.The scaled-image comparison is presented in Table 4 and Figure 3.
Fisheye distortion
The reported results identify SIFT as strongest in one tested scenario, while noisy images favor nearly equal performance from SIFT and ORB. Noise can affect keypoints because it is randomly distributed.
- SIFT achieved the highest matching rate in the reported scenario, with fewer correct matches than in previous scenarios.
- Noisy images: 30% salt-and-pepper noise produced the best matching rates for SIFT and ORB, with almost equal performance.
- Noisy images: Randomly distributed noise may affect some keypoints, yet SIFT and ORB retained nearly equal matching performance.
IV. CONCLUSION
The paper evaluates SIFT, SURF, and ORB under multiple image transformations using matching rate, keypoint counts, and execution time. ORB is fastest overall, while SIFT performs best in most scenarios, with exceptions for near-90-degree rotations and near-equal noisy-image performance.
- The study applies scaling, rotation, noise, fisheye distortion, and shearing, evaluating keypoint counts, matching rates, and execution times.
- ORB is the fastest algorithm, while SIFT performs best in most tested scenarios.
- For rotations proportional to 90 degrees, ORB and SURF outperform SIFT; on noisy images, ORB and SIFT perform almost similarly.