Log-Polar Space Convolution for Convolutional Neural Networks

Bing Su, Ji-Rong WenarXiv 2021 · arXiv

Introduction

In computer vision, the "Local Receptive Field" (LRF) is fundamental: to understand a specific pixel, a model must consider its surrounding context. However, standard Convolutional Neural Networks (CNNs) face a "quadratic trap." As the kernel size increases to capture more context, the number of parameters grows quadratically (e.g., a $15 \times 15$ kernel requires 225 parameters per channel). While techniques like dilated convolutions or stacking small layers exist, they often introduce "holes" in the data or suffer from optimization issues like vanishing gradients.

This paper introduces Log-Polar Space Convolution (LPSC), a method that breaks this trap by mimicking the foveated vision of the human eye. By rethinking the geometry of the local neighborhood, LPSC allows for exponentially larger receptive fields with minimal parameter overhead.

Overview of LPSC: (a) foveated receptive fields applied to an image, and (b) the structured log-polar grid division.

Methodology: The LPSC Kernel

The core innovation of LPSC is the division of the local receptive field into bins based on log-polar coordinates rather than a Cartesian grid.

  1. Log-Polar Division: The area around a center pixel is divided into regions (bins) that grow exponentially larger as they move away from the focal point.
  2. Weight Sharing: All pixels within the same bin share a single weight. This structural prior assumes that high resolution is only necessary at the center, while peripheral information can be aggregated. For instance, an $11 \times 11$ area might be represented by only ~20 parameters.
  3. Elliptical Generalization: Unlike standard circular log-polar transforms, LPSC can utilize elliptical kernels. This allows the model to adapt to different orientations and aspect ratios within the spatial context.
The LPSC kernel generalized to an elliptical form, defined by eccentricity $e$ and orientation $\alpha$.
  1. Implementation: To leverage existing optimized deep learning frameworks, the authors use a "log-polar space pooling" approach. The input is transformed into a map where standard convolutions can simulate the LPSC effect. A separate $1 \times 1$ convolution is typically maintained for the center pixel to preserve fine-grained focal details.

Key Insights and Bio-mimicry

LPSC is deeply rooted in biological principles. The human retina features a high-resolution fovea and a lower-resolution periphery. By adopting this "foveated" approach, LPSC achieves:

  • Exponential Growth: Receptive fields expand much faster than standard kernels, providing "long-range vision" efficiently.
  • Scale Robustness: The logarithmic mapping makes the network less sensitive to object size; small objects are effectively "enlarged" in the coordinate space, while large objects are "compressed."
  • Structured Priors: Unlike dilated convolutions that sample discrete pixels, LPSC maintains a continuous, geometric spatial prior.

Experimental Results

The authors validated LPSC across multiple domains:

  • Classification: On CIFAR-100, an LPSC-enhanced AlexNet achieved 47.31% accuracy (vs. 43.87% for the baseline) with fewer parameters.
  • Efficiency: LPSC significantly reduces computational cost. In one test, LPSC required 11.42M FLOPs, compared to 24.71M for dilated convolutions.
  • Semantic Segmentation: Integrating LPSC into DeepLabv3+ improved Mean Intersection over Union (mIoU) on PASCAL VOC by 1.1%, demonstrating the value of increased spatial context.
  • Medical Imaging: LPSC reached state-of-the-art performance on the DRIVE dataset for retinal vessel detection, proving adept at capturing long, branching structures.

Conclusion and Future Directions

Log-Polar Space Convolution offers a compelling alternative to traditional kernels by providing a biologically inspired, parameter-efficient way to expand receptive fields. While currently implemented via a pooling workaround, future performance gains are expected with the development of native CUDA kernels for direct log-polar indexing. Additionally, optimizing the hyperparameters governing distance and direction levels remains an area for further exploration.

Summary updated

Abstract

Convolutional neural networks use regular quadrilateral convolution kernels to extract features. Since the number of parameters increases quadratically with the size of the convolution kernel, many popular models use small convolution kernels, resulting in small local receptive fields in lower layers. This paper proposes a novel log-polar space convolution (LPSC) method, where the convolution kernel is elliptical and adaptively divides its local receptive field into different regions according to the relative directions and logarithmic distances. The local receptive field grows exponentially with the number of distance levels. Therefore, the proposed LPSC not only naturally encodes local spatial structures, but also greatly increases the single-layer receptive field while maintaining the number of parameters. We show that LPSC can be implemented with conventional convolution via log-polar space pooling and can be applied in any network architecture to replace conventional convolutions. Experiments on different tasks and datasets demonstrate the effectiveness of the proposed LPSC. Code is available at https://github.com/BingSu12/Log-Polar-Space-Convolution.
BibTeX
@misc{su2021log,
  author = {Bing Su and Ji-Rong Wen},
  title = {{Log-Polar Space Convolution for Convolutional Neural Networks}},
  year = {2021},
  eprint = {2107.11943},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.11943},
}
Figure