Fast 2-D Complex Gabor Filter with Kernel Decomposition

Suhyuk Um, Jaeyoon Kim, Dongbo Min

Abstract

2-D complex Gabor filtering has found numerous applications in the fields of computer vision and image processing. Especially, in some applications, it is often needed to compute 2-D complex Gabor filter bank consisting of the 2-D complex Gabor filtering outputs at multiple orientations and frequencies. Although several approaches for fast 2-D complex Gabor filtering have been proposed, they primarily focus on reducing the runtime of performing the 2-D complex Gabor filtering once at specific orientation and frequency. To obtain the 2-D complex Gabor filter bank output, existing methods are repeatedly applied with respect to multiple orientations and frequencies. In this paper, we propose a novel approach that efficiently computes the 2-D complex Gabor filter bank by reducing the computational redundancy that arises when performing the Gabor filtering at multiple orientations and frequencies. The proposed method first decomposes the Gabor basis kernels to allow a fast convolution with the Gaussian kernel in a separable manner. This enables reducing the runtime of the 2-D complex Gabor filter bank by reusing intermediate results of the 2-D complex Gabor filtering computed at a specific orientation. Furthermore, we extend this idea into 2-D localized sliding discrete Fourier transform (SDFT) using the Gaussian kernel in the DFT computation, which lends a spatial localization ability as in the 2-D complex Gabor filter. Experimental results demonstrate that our method runs faster than state-of-the-arts methods for fast 2-D complex Gabor filtering, while maintaining similar filtering quality.

Introduction

Gabor filters are essential components in early-stage computer vision, mimicking the human visual system's ability to extract features at specific orientations and scales. They are widely used for texture analysis, fingerprint recognition, and face detection. However, practical applications typically require a filter bank—a collection of many filters (often 40 or more) covering various orientations and frequencies to capture a complete feature set.

Figure 1: A 2-D complex Gabor filter bank consisting of multiple orientations and frequencies.

Performing 2-D convolutions for dozens of different kernels is computationally expensive. While previous "fast" methods optimized single filter applications, they often ignored the massive computational redundancy occurring when running multiple similar filters on the same image. This paper addresses this redundancy through kernel decomposition and result recycling.

Methodology

The proposed approach reduces computation time through three primary strategies:

Kernel Decomposition and Separability

The 2-D Gabor kernel is decomposed into 1-D horizontal and vertical components. By leveraging trigonometric identities, the complex exponential term is broken down so that the real and imaginary parts share the same Gaussian smoothing operations. A fast recursive Gaussian filter serves as the computational backbone for these operations.

Symmetry-Based Recycling

A key observation in this work is that for a fixed frequency, horizontal filtering results for an orientation $\theta$ and its symmetric counterpart $\pi - \theta$ are complex conjugates. By computing the horizontal pass once and "recycling" the result via conjugation for the symmetric orientation, the algorithm significantly reduces redundant operations across the filter bank.

Figure 2: (a) Log polar grid and (b) Rectangular grid used for sampling orientations and frequencies, highlighting the symmetry used for recycling results.

Localized Sliding DFT (SDFT)

The decomposition logic is extended to the Sliding Discrete Fourier Transform. Unlike standard SDFTs that use a "box" window (which can cause artifacts), this version employs a Gaussian window for superior spatial localization, optimized via the same kernel decomposition technique.

Experimental Results

The method was evaluated using the USC-SIPI image database, covering aerial shots, standard portraits, video sequences, and textures. Key findings include:

  • Execution Time: For a filter bank with 40 coefficients (8 orientations, 5 frequencies), the proposed method was approximately 40% faster than standard recursive Gabor filters.
  • Computational Efficiency: In a bank of 30 orientations, the method required only 30 multiplications per pixel, compared to 52 for standard recursive approaches and 34 for previous IIR methods.
  • SDFT Performance: The localized SDFT demonstrated gains up to 3x faster than traditional IIR-based implementations as window sizes increased.
  • Quality: Despite using IIR approximations for the Gaussian, the Signal-to-Error Ratio (SER) remained high enough to ensure negligible impact on practical computer vision performance.

Conclusion

This paper presents a novel approach to 2-D complex Gabor filtering that prioritizes the efficiency of the entire filter bank rather than individual filters. By exploiting mathematical symmetries and kernel separability, the method achieves state-of-the-art speeds without sacrificing filtering quality. Future work will explore mapping these decompositions to parallel architectures like GPUs and applying the Gaussian-windowed SDFT to more complex vision pipelines.

Summary updated

Abstract

2-D complex Gabor filtering has found numerous applications in the fields of computer vision and image processing. Especially, in some applications, it is often needed to compute 2-D complex Gabor filter bank consisting of the 2-D complex Gabor filtering outputs at multiple orientations and frequencies. Although several approaches for fast 2-D complex Gabor filtering have been proposed, they primarily focus on reducing the runtime of performing the 2-D complex Gabor filtering once at specific orientation and frequency. To obtain the 2-D complex Gabor filter bank output, existing methods are repeatedly applied with respect to multiple orientations and frequencies. In this paper, we propose a novel approach that efficiently computes the 2-D complex Gabor filter bank by reducing the computational redundancy that arises when performing the Gabor filtering at multiple orientations and frequencies. The proposed method first decomposes the Gabor basis kernels to allow a fast convolution with the Gaussian kernel in a separable manner. This enables reducing the runtime of the 2-D complex Gabor filter bank by reusing intermediate results of the 2-D complex Gabor filtering computed at a specific orientation. Furthermore, we extend this idea into 2-D localized sliding discrete Fourier transform (SDFT) using the Gaussian kernel in the DFT computation, which lends a spatial localization ability as in the 2-D complex Gabor filter. Experimental results demonstrate that our method runs faster than state-of-the-arts methods for fast 2-D complex Gabor filtering, while maintaining similar filtering quality.
BibTeX
@article{um2018fast,
  author = {Suhyuk Um and Jaeyoon Kim and Dongbo Min},
  title = {{Fast 2-D Complex Gabor Filter with Kernel Decomposition}},
  journal = {IEEE Trans. Image Process.},
  year = {2018},
  url = {https://openreview.net/forum?id=oQu37X4k0Z},
}
Figure