Abstract
In this paper, we present an application of wavelet analysis on feature points matching. Wavelet coefficients produced by wavelet transform can best display the local character of an image. Through a new feature description and matching method, we build a feature point database for each image to improve matching efficiency. Experimental results demonstrate that this method is practical, feasible, precise, and fast.
Introduction
Feature point matching is a fundamental task in computer vision, essential for applications like stereo vision and object tracking. The process typically involves detection, description, and matching. While traditional tools like the Harris corner detector are effective at finding interest points, they often lack scale robustness. Conversely, high-dimensional descriptors like SIFT provide robustness but at a significant computational cost. This paper proposes a bridge: using Wavelet Analysis to create compact, statistical descriptors that are both efficient and discriminative.
Methodology
The proposed pipeline utilizes multiresolution analysis to decompose image patches into smooth and detailed components. The process consists of four main stages:
- Detection: A Harris corner detector identifies initial points of interest.
- Patch Extraction: For every interest point, a small sub-image (typically $7 \times 7$ or $9 \times 9$ pixels) is cropped around the center.
- Wavelet Statistical Description: A Daubechies wavelet transform (db-3 at level 2) is applied to these patches. Rather than using raw coefficients, the algorithm extracts nine specific statistics: Mean, Median, Mode, Maximum, Minimum, Range, Standard Deviation (SD), Median Absolute Deviation (MdAD), and Mean Absolute Deviation (MnAD).
- Voting-Based Matching: Matching is performed via a probability-based voting system. For each statistical category, the algorithm identifies the candidate point in the target image with the minimum difference. A point is considered an exact match if it "wins" in more than 87.5% of the categories, and an approximate match if it wins in 62.5% to 87.5%.
Feature Point Database
The core innovation lies in reducing complex wavelet data into a searchable "Image Feature Table." By compressing the transform into nine statistical fingerprints, the system drastically reduces the data storage and comparison overhead. As shown in the feature database, these statistics vary significantly between different interest points, providing the necessary discriminatory power to prevent false positives during the matching phase.
Experimental Results
The algorithm was tested on custom image pairs, including facial images with variations in pose. Qualitative results show dense and accurate point localization across different views. The efficiency of the method is attributed to comparing 9-dimensional statistical vectors rather than the high-dimensional gradient histograms used in traditional methods like SIFT. The voting system effectively acts as a "committee of experts," ensuring that a match is robust across multiple independent measures.
Conclusion and Future Work
This wavelet-based approach offers a computationally light alternative for feature matching. While visual results are promising, future work should include quantitative benchmarking against standard datasets (like COCO) and an evaluation of the algorithm's performance under extreme scale and rotation changes. Additionally, exploring larger patch windows or different wavelet families could further enhance robustness in noisy environments.
Summary updated