An efficient multi-resolution variational Retinex scheme for the radiometric correction of airborne remote sensing images

Huifang Li, Xiaojing Wang, Huanfeng Shen, Qiangqiang Yuan, Liangpei ZhangInternational Journal of Remote Sensing 2016 · DOI

TL;DR

This paper introduces a novel multi-resolution variational Retinex (MVR) method to efficiently correct non-uniform brightness in airborne remote sensing images. By combining a Gaussian pyramid for multi-resolution processing with the fast Split Bregman (SB) iteration for optimization, the method significantly speeds up computation while preserving spectral information, achieving up to 110x faster results than traditional single-resolution approaches.

What Problem Does This Paper Address?

Airborne remote sensing images often suffer from non-uniform brightness, a degradation caused by factors like lens vignetting, viewing angles, and illumination variations. This unevenness significantly reduces the precision of image interpretation and can negatively impact subsequent applications such as image mosaicking, feature extraction, and target recognition. Existing methods, while capable of correcting this, often face challenges with computational efficiency, especially for large remote sensing datasets, and can sometimes lead to a loss of spectral information, resulting in less natural-looking images. This paper aims to address these limitations by proposing a more efficient and accurate radiometric correction technique.

Methods

The core of the proposed method is a Multi-Resolution Variational Retinex (MVR) scheme. This approach leverages a Gaussian pyramid to decompose the image into multiple resolution levels. The variational Retinex model, which is based on the Retinex theory (observed image = illumination * reflectance), is then applied iteratively, starting from the coarsest resolution and progressively refining the solution at finer resolutions.

To tackle the computational burden of the variational model, the paper incorporates the fast Split Bregman (SB) iteration method. This optimization technique is particularly effective for L1-regularized problems, which are common in image processing tasks like total variation denoising. The SB iteration breaks down the complex optimization problem into simpler, iterative sub-problems, leading to significantly faster convergence. The overall process involves generating the Gaussian pyramid, optimizing the illumination component at each level (starting from the coarsest), and then using the results to initialize and optimize at the next finer level, ultimately yielding the corrected reflectance component.

Datasets

The paper evaluates the proposed method using both synthetic and real remote sensing image data.

  • Synthetic Data: A standard 512x512 single-band image was used as the original, uniformly illuminated image. Two types of degraded images were generated: one with linear brightness change along the horizontal direction and another with brightness change following a Gaussian kernel spreading direction.
  • Real Aerial Data: Three aerial images were used:
  • A 600x400 pixel, three-band (red, green, blue) image.
  • A 1000x1000 pixel, three-band image.
  • A 950x600 pixel, three-band image.

Key Insights and Ideas

  • Multi-resolution as an Accelerator: The Gaussian pyramid isn't just for decomposition; it's a crucial acceleration strategy. By solving the problem at coarser levels first, the method obtains a good initial estimate for finer levels, drastically reducing the number of iterations needed at each stage.
  • Split Bregman for Efficiency: The choice of the Split Bregman (SB) iteration is key. It's a powerful tool for L1-regularized problems and efficiently handles the complex optimization required by the variational Retinex model, making it much faster than traditional gradient descent methods.
  • Color Preservation is Not Sacrificed: A significant finding is that the multi-resolution approach, by producing smoother illumination components, actually improves color preservation compared to single-resolution methods, leading to more visually natural results.
  • Optimal Layer Number: The study demonstrates that there's an optimal number of layers (found to be 4 in their experiments) for the Gaussian pyramid, balancing computational efficiency with visual quality. Too few layers might not capture enough detail, while too many can lead to diminishing returns and increased computation.

Key Results

  • Significant Speedup: The proposed MVRSB method achieved substantial speed improvements. Compared to the traditional single-resolution steepest descent (SVRSD) method, MVRSB was found to be 60-110 times faster on synthetic data and up to 80% faster on aerial data. In some cases, it was one-sixtieth the time of SVRSD.
  • Effective Brightness Correction: Visual evaluations and quantitative metrics (PSNR, RMSE, SSIM, DGA) showed that the MVRSB method effectively corrected non-uniform brightness in both synthetic and real aerial images.
  • Improved Spectral Information: The multi-resolution approach, particularly with 4 layers, resulted in more natural-looking colors and better preservation of spectral information compared to single-resolution methods.
  • Robustness: The method was found to be robust for aerial data covering urban areas, with consistent parameter settings yielding good results.

Limitations and Future Work

  • Layer Number Optimization: While the paper identifies an optimal layer number (4) based on their experiments, the authors suggest that further exploration into layer number determination for different types of images could be beneficial.
  • Real-World Scene Complexity: The paper notes that in real scenes, brightness distribution might not always follow simple mathematical formulas. While the method is robust, future work could explore more adaptive approaches for highly complex or unusual illumination patterns.
  • Large-Scale Data Processing: The authors explicitly state that the efficiency gains make their method suitable for large-scale remote sensing data and suggest this as a direction for future research.

Summary updated

Abstract

The brightness non-uniformity caused by vignetting effects, viewing, and illumination angels in remote sensing images reduces the interpretation precision. A multi-resolution variational Retinex scheme is proposed in this paper to efficiently correct the non-uniform brightness in airborne remote sensing images. This variational Retinex model is non-linear, constrained by the grey-world assumption and the total variation regularization. A Gaussian image pyramid is used to construct the multi-resolution scheme. The multi-resolution scheme reduces the calculation burden and raises the calculation efficiency. The fast split Bregman (SB) iteration method is employed to optimize the proposed non-linear model in each level of the multi-resolution scheme. This decomposes the complicated model into several simple sub-problems and greatly improves the calculation efficiency. The multi-resolution scheme embedded with the SB iteration method was applied to both synthetic and real remote sensing images. The experimental results show that the brightness non-uniformity can be corrected, and the spectral information can be effectively restored. Moreover, the calculation efficiency is raised by about 60–110 times, compared to the traditional single-resolution solving method.
BibTeX
@article{li2016efficient,
  author = {Huifang Li and Xiaojing Wang and Huanfeng Shen and Qiangqiang Yuan and Liangpei Zhang},
  title = {{An efficient multi-resolution variational Retinex scheme for the radiometric correction of airborne remote sensing images}},
  journal = {International Journal of Remote Sensing},
  year = {2016},
  doi = {10.1080/01431161.2016.1145364},
}
Figure