Entropy-based Adaptive Image Denoising

Long Ye, Haijun Gao, Qin ZhangProceedings of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013) 2013 · DOI

Abstract

Image entropy, which not only describes the average amount of information about the image source, but also reflects the statistical characteristics of image data, can be described as the properties of image features and image processing basis. This paper is aimed to put forward an adaptive image denoising solution by setting the threshold and analyzing the impact of the original image data, which comes from the image noise under different entropy. For the blocks in image with low entropy value, we apply Haar wavelet method to denoise; for the medium entropy, we use field of expert (FoE) model as the prior information of the medium region denoising to achieve the image denoising of these blocks, and for the high entropy blocks, because little difference existed between ideal and noised image in human perception, no processing is implemented in this kind of blocks. The adaptability of our framework is embodied in the adjusting of the thresholds to classify different entropy blocks. Experiment results demonstrate the advantages of our framework measured by both PSNR and SSIM.

Introduction

Image denoising remains a fundamental challenge in computer vision, requiring a delicate balance between noise suppression and detail preservation. Traditional methods often apply a uniform algorithm across an entire image. However, smooth regions and complex textures have vastly different statistical properties. Simple transformation methods like wavelets are computationally efficient and effective for smooth areas but often fail to reconstruct intricate textures. Conversely, sophisticated "learned" models like the Fields of Experts (FoE) framework can reconstruct details but are computationally expensive and prone to artifacts if the priors do not perfectly match the image content.

A key observation driving this research is perceptual masking: the human visual system is significantly less sensitive to noise in highly textured, high-entropy regions than in smooth, low-entropy areas. Treating an entire image with a heavy-handed denoising algorithm can lead to over-smoothing or unnecessary computational overhead.

Entropy-Based Adaptive Framework

The proposed method introduces a "triage" approach to denoising, classifying image regions by their local complexity. The process begins by dividing the noisy image into non-overlapping $8 \times 8$ blocks and calculating the entropy for each.

Comparison of original and noisy image blocks across high, medium, and low entropy levels.

Based on two optimized thresholds ($\sigma_1$ and $\sigma_2$), blocks are processed through one of three specialized pipelines:

  1. Low Entropy ($H_i < \sigma_2$): These smooth regions are processed using Haar Wavelet denoising. This method is efficient and well-suited for preserving basic structures in simple areas.
  2. Medium Entropy ($\sigma_2 \le H_i \le \sigma_1$): For regions with moderate detail, the Fields of Experts (FoE) model is employed. This sophisticated framework uses learned image priors to accurately reconstruct signals where noise is visible but underlying structure remains.
  3. High Entropy ($H_i > \sigma_1$): In highly complex textures, the authors implement no processing. Because the visual difference between the ideal and noisy signal is minimal in these regions, skipping denoising preserves the original texture and saves computational resources.

The thresholds $\sigma_1$ and $\sigma_2$ are iteratively adjusted to maximize performance metrics like Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index (SSIM).

Experimental Results

The framework was evaluated using standard test images like "Lena" and the "Ivy Wall" sequence, which provides a natural gradient of entropy levels. Additive zero-mean Gaussian noise was applied at various standard deviations ($\sigma = 10, 15, 20, 25$).

The hybrid approach consistently outperformed standalone Haar and FoE methods. For the "Lena" image with a noise variance of 10, the proposed method achieved a PSNR of 29.74, compared to 29.66 for Haar and 29.70 for FoE. More importantly, the method showed a clear advantage in SSIM, scoring 0.9406 against Haar's 0.9319 and FoE's 0.9401. Qualitatively, the adaptive approach avoids the "smearing" effect sometimes seen when applying FoE to low-entropy regions, maintaining better edge and texture fidelity.

Conclusion

By leveraging local entropy as a guide, this adaptive framework focuses computational power where it is most effective. The decision to utilize perceptual masking for high-entropy regions prevents over-smoothing and improves overall visual quality. Future work will explore more advanced denoising techniques for low-entropy regions and investigate offline prior integration to reduce the computational burden of the FoE model training.

Summary updated

Abstract

Image entropy, which not only describes the average amount of information about the image source, but also reflects the statistical characteristics of image data, can be described as the properties of image features and image processing basis. This paper is aimed to put forward an adaptive image denoising solution by setting the threshold and analyzing the impact of the original image data, which comes from the image noise under different entropy. For the blocks in image with low entropy value, we apply Haar wavelet method to denoise; for the medium entropy, we use field of expert (FoE) model as the prior information of the medium region denoising to achieve the image denoising of these blocks, and for the high entropy blocks, because little difference existed between ideal and noised image in human perception, no processing is implemented in this kind of blocks. The adaptability of our framework is embodied in the adjusting of the thresholds to classify different entropy blocks. Experiment results demonstrate the advantages of our framework measured by both PSNR and SSIM.
BibTeX
@inproceedings{ye2013entropy,
  author = {Long Ye and Haijun Gao and Qin Zhang},
  title = {{Entropy-based Adaptive Image Denoising}},
  booktitle = {Proceedings of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013)},
  year = {2013},
  doi = {10.2991/iccsee.2013.81},
}
Figure