Invariant Pattern Recognition with Log-Polar Transform and Dual-Tree Complex Wavelet-Fourier Features

Guangyi Chen, Adam KrzyżakSensors 2023 · DOI

TL;DR

This paper introduces a novel 2D pattern recognition method that achieves invariance to translation, rotation, and scaling by combining log-polar transform, Dual-Tree Complex Wavelet Transform (DTCWT), and Fast Fourier Transform (FFT). The method leverages intermediate-resolution DTCWT sub-bands to effectively capture features while mitigating noise, outperforming existing approaches on benchmark datasets.

What Problem Does This Paper Address?

In real-world applications like optical character recognition (OCR), face recognition, and biometrics, recognizing patterns accurately despite variations in their position, orientation, and size is crucial. Traditional pattern recognition methods often struggle with these transformations, leading to poor performance. Achieving invariance to translation, rotation, and scaling simultaneously, especially in a multiresolution manner that balances feature preservation and noise reduction, remains a significant challenge.

Methods

The proposed method combines three key techniques:

  • Log-Polar Transform: This transform converts Cartesian coordinates to a log-polar system. Crucially, it maps rotation and scaling in the original image to simple spatial shifts in the log-polar domain. This is a standard technique for achieving rotation and scale invariance.
  • Dual-Tree Complex Wavelet Transform (DTCWT): DTCWT decomposes an image into multiple sub-bands at different resolutions. It offers a degree of translation invariance and, importantly, allows for the selection of specific resolution levels. The authors highlight that very low-resolution sub-bands lose fine details, while very high-resolution sub-bands are prone to noise. Therefore, they focus on intermediate-resolution sub-bands (specifically the 3rd, 4th, and 5th) as they are considered optimal for capturing robust invariant features.
  • 2D Fast Fourier Transform (FFT2): Applying FFT2 to the DTCWT sub-bands further enhances translation invariance. The magnitude of the Fourier coefficients is invariant to spatial shifts, making it a powerful tool for feature extraction in this context.

The overall pipeline involves:

  1. Translating the input image to its centroid.
  2. Converting the image to log-polar coordinates (128x128 pixels).
  3. Applying DTCWT for 5 decomposition scales.
  4. Constructing complex sub-bands by combining the two trees of DTCWT.
  5. Applying FFT2 to each complex sub-band and extracting their spectra.
  6. Using the 3rd, 4th, and 5th sub-band coefficients for recognition via a nearest neighbor classifier.

Datasets

The paper evaluates the proposed method on two datasets:

  • Printed Chinese Character Dataset: A binary dataset containing 85 characters.
  • 2D Aircraft Dataset: A binary dataset containing 20 aircraft images.

These datasets were subjected to various deformations, including:

  • Scaling factors: Ranging from 0.2 to 1.0.
  • Rotation angles: Including 30°, 60°, 90°, 120°, 150°, 180°, 210°, 240°, and 270°.
  • Noise levels: Introduced using Signal-to-Noise Ratio (SNR) values from 20 down to 0.5.

Key Insights and Ideas

  • The Power of Intermediate Resolutions: The core insight is that not all wavelet sub-bands are created equal for invariant pattern recognition. By strategically selecting intermediate-resolution sub-bands from the DTCWT, the method effectively balances the need to retain crucial pattern features with the necessity of discarding noise present in higher-resolution bands.
  • Synergistic Combination: The paper demonstrates that the combination of log-polar transform (for rotation/scaling), DTCWT (for multiresolution analysis and translation invariance), and FFT2 (for enhanced translation invariance) creates a powerful feature extraction pipeline. Each component addresses specific invariance requirements, and their integration yields superior results.
  • Computational Efficiency: The authors claim their method is competitive with Convolutional Neural Networks (CNNs) in terms of speed, as it avoids the extensive training time typically required for CNNs.

Key Results

The proposed method consistently outperformed two baseline methods (log-polar-FFT2 and log-polar DWT-FFT2) across various experimental conditions.

  • Superior Recognition Rates: In most testing scenarios involving combinations of rotation angles, scaling factors, and noise levels, the proposed method achieved higher correct recognition rates on both the Chinese character and aircraft datasets. For instance, in many cases, the proposed method achieved 100% recognition rates under moderate to high SNR levels and for larger scaling factors, while the other methods showed a decline.
  • Robustness to Noise: The method demonstrated good robustness to noise, particularly at higher SNRs. While performance degraded at very low SNRs (e.g., 0.5), it still often outperformed the baselines.
  • Speed: The paper states the method is fast in terms of CPU computational time for invariant pattern recognition.

Limitations and Future Work

  • Performance at Very Low SNR: The paper acknowledges that performance degrades significantly at very low SNR levels (e.g., 0.5), where recognition rates can drop considerably.
  • Future Directions:
    • Image Denoising Preprocessing: The authors suggest incorporating image denoising techniques (referencing their prior work) as a preprocessing step to further improve recognition results, especially in noisy conditions.
  • Deep Learning Integration: They also plan to explore Deep Convolutional Neural Networks (DNNs) for invariant pattern recognition, recognizing their success in real-world applications.

Summary updated

Abstract

In this paper, we propose a novel method for 2D pattern recognition by extracting features with the log-polar transform, the dual-tree complex wavelet transform (DTCWT), and the 2D fast Fourier transform (FFT2). Our new method is invariant to translation, rotation, and scaling of the input 2D pattern images in a multiresolution way, which is very important for invariant pattern recognition. We know that very low-resolution sub-bands lose important features in the pattern images, and very high-resolution sub-bands contain significant amounts of noise. Therefore, intermediate-resolution sub-bands are good for invariant pattern recognition. Experiments on one printed Chinese character dataset and one 2D aircraft dataset show that our new method is better than two existing methods for a combination of rotation angles, scaling factors, and different noise levels in the input pattern images in most testing cases.
BibTeX
@inproceedings{chen2023invariant,
  author = {Guangyi Chen and Adam Krzyżak},
  title = {{Invariant Pattern Recognition with Log-Polar Transform and Dual-Tree Complex Wavelet-Fourier Features}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23083842},
}
Figure