On Image segmentation using Fractional Gradients-Learning Model Parameters using Approximate Marginal Inference

Anish Acharya, Uddipan Mukherjee, Charless C. FowlkesarXiv 2016 · arXiv

Abstract

This paper presents a novel framework for image segmentation that utilizes fractional-order gradients to optimize the balance between edge sharpness and noise suppression. By replacing traditional integer-order derivatives with a tunable fractional derivative of Gaussian (FDOG), the proposed method achieves a 79.2% average accuracy on the Stanford Backgrounds Dataset, outperforming standard first and second-order derivative approaches.

Introduction

In computer vision, gradients are fundamental to edge detection and object recognition. However, traditional methods face a persistent trade-off: low-order derivatives (e.g., 1st order) are robust to noise but produce blurry edges, while high-order derivatives (e.g., 2nd order) offer sharp responses but are highly sensitive to noise. This research explores the "in-between" space using fractional calculus, allowing the gradient order to be a continuous parameter optimized for natural scene statistics.

Methodology

The core of the approach is the Fractional Derivative of Gaussian (FDOG). By employing discrete approximations of $n$-th order differentiation, the authors create a family of filters where the order $v$ is a tunable parameter.

Optimal Order Selection and Feature Engineering

The researchers developed a custom score function to balance the Peak Signal-to-Noise Ratio (PSNR) against the False Discovery Rate (DE). Their findings indicate that an order of $v=0.6$ to $0.7$ typically provides the optimal balance for natural images. These filters are then used to construct Fractional HOG (FHOG) features, which serve as the input for a segmentation model.

The FDOG approach produces sharp, noise-robust edge maps compared to traditional integer-order filters.

Segmentation Model

The system utilizes a Conditional Random Field (CRF) to label pixels. Training is conducted using Tree Re-weighted (TRW) Belief Propagation and a specialized "Clique Loss" function. This differentiable pipeline allows for efficient parameter optimization, enabling the model to learn complex spatial relationships between neighboring pixels more effectively than standard maximum likelihood methods.

Experimental Results

The model was evaluated on the Stanford Backgrounds Dataset (715 outdoor images), the Berkeley Segmentation Dataset (BSDS500), and the CSIQ Image Database for noise robustness tests.

Qualitative results on the Stanford Backgrounds Dataset: Original image (left), ground truth (center), and predicted segmentation (right).

Key results include:

  • Accuracy: The FDOG-based model achieved a 79.2% average accuracy on the Stanford Backgrounds Dataset.
  • Error Reduction: This represents a nearly 10% reduction in error compared to the previous state-of-the-art framework using standard gradients.
  • Robustness: The FDOG filters naturally suppress noise while maintaining structural detail, even in images corrupted with white noise.

Discussion and Conclusion

The primary insight of this work is that the optimal derivative order for natural images is non-integer. Fractional orders introduce "hyper-damped" poles into the mathematical model, providing the flexibility needed to capture complex dynamics that integer orders miss. While the 20.8% remaining error on the Stanford dataset indicates room for improvement, the qualitative superiority of FDOG edge maps suggests that fractional calculus is a powerful tool for low-level vision tasks. Future work will explore extending these concepts into the frequency domain and integrating texture-based features.

Summary updated

Abstract

Estimates of image gradients play a ubiquitous role in image segmentation and classification problems since gradients directly relate to the boundaries or the edges of a scene. This paper proposes an unified approach to gradient estimation based on fractional calculus that is computationally cheap and readily applicable to any existing algorithm that relies on image gradients. We show experiments on edge detection and image segmentation on the Stanford Backgrounds Dataset where these improved local gradients outperforms state of the art, achieving a performance of 79.2% average accuracy.
BibTeX
@misc{acharya2016image,
  author = {Anish Acharya and Uddipan Mukherjee and Charless C. Fowlkes},
  title = {{On Image segmentation using Fractional Gradients-Learning Model Parameters using Approximate Marginal Inference}},
  year = {2016},
  eprint = {1605.02240},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1605.02240},
}
Figure