Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Sharpness Based Loss Function for Removing Out-of-Focus Blur

Uditangshu Aurangabadkar, Darren Ramsook, Anil Kokaram

TL;DR

This work addresses removing out-of-focus blur by introducing a differentiable, no-reference sharpness metric $Q$ as a loss in an encoder–decoder deblurring network. It pairs this loss with a new high-resolution dataset of real-world defocused images and a two-stage training strategy, yielding sharper restorations and improved perceptual quality (LPIPS) alongside competitive PSNR and SSIM. The method demonstrates notable gains on synthetic defocus benchmarks and real-world images, and it shows promise for guiding state-of-the-art models toward crisper outputs without sacrificing fidelity. The work provides a practical approach to integrating perceptual sharpness into training and offers a valuable dataset for evaluation in real-camera blur scenarios.

Abstract

The success of modern Deep Neural Network (DNN) approaches can be attributed to the use of complex optimization criteria beyond standard losses such as mean absolute error (MAE) or mean squared error (MSE). In this work, we propose a novel method of utilising a no-reference sharpness metric Q introduced by Zhu and Milanfar for removing out-of-focus blur from images. We also introduce a novel dataset of real-world out-of-focus images for assessing restoration models. Our fine-tuned method produces images with a 7.5 % increase in perceptual quality (LPIPS) as compared to a standard model trained only on MAE. Furthermore, we observe a 6.7 % increase in Q (reflecting sharper restorations) and 7.25 % increase in PSNR over most state-of-the-art (SOTA) methods.

A Sharpness Based Loss Function for Removing Out-of-Focus Blur

TL;DR

This work addresses removing out-of-focus blur by introducing a differentiable, no-reference sharpness metric as a loss in an encoder–decoder deblurring network. It pairs this loss with a new high-resolution dataset of real-world defocused images and a two-stage training strategy, yielding sharper restorations and improved perceptual quality (LPIPS) alongside competitive PSNR and SSIM. The method demonstrates notable gains on synthetic defocus benchmarks and real-world images, and it shows promise for guiding state-of-the-art models toward crisper outputs without sacrificing fidelity. The work provides a practical approach to integrating perceptual sharpness into training and offers a valuable dataset for evaluation in real-camera blur scenarios.

Abstract

The success of modern Deep Neural Network (DNN) approaches can be attributed to the use of complex optimization criteria beyond standard losses such as mean absolute error (MAE) or mean squared error (MSE). In this work, we propose a novel method of utilising a no-reference sharpness metric Q introduced by Zhu and Milanfar for removing out-of-focus blur from images. We also introduce a novel dataset of real-world out-of-focus images for assessing restoration models. Our fine-tuned method produces images with a 7.5 % increase in perceptual quality (LPIPS) as compared to a standard model trained only on MAE. Furthermore, we observe a 6.7 % increase in Q (reflecting sharper restorations) and 7.25 % increase in PSNR over most state-of-the-art (SOTA) methods.
Paper Structure (9 sections, 7 equations, 7 figures, 3 tables)

This paper contains 9 sections, 7 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset
  4. Network Architecture
  5. Training
  6. Experimental Results
  7. Synthetic Dataset
  8. Real Dataset
  9. Conclusion

Figures (7)

  • Figure 1: Crop of an image from the proposed dataset alongside three out-of-focus counterparts.
  • Figure 2: Proposed network architecture. Yellow, green and pink panels represent the encoder, decoder and bottleneck block. Red arrows represent the skip connections.
  • Figure 3: The visual effect of a gradual increase in $\beta$ used during fine-tuning. As $\beta$ increases, the images produced are sharper (N.B. the texture around the bricks)
  • Figure 4: Comparison of methods for removing synthetic out-of-focus blur. Our fine-tuned method produces a sharper edge around the roof.
  • Figure 5: Comparison between sharpness ($Q$) and perceptual quality (SSIM). Our methods (highlighted in red) generate images with higher structural similarity, as well as sharpness.
  • ...and 2 more figures