Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Study in Dataset Pruning for Image Super-Resolution

Brian B. Moser, Federico Raue, Andreas Dengel

TL;DR

This paper tackles the data inefficiency of image SR by proposing a loss-value–based core-set pruning strategy. Using a simple pre-trained SRCNN to estimate reconstruction loss, it forms a 50% core-set by selecting high-loss samples and further improves performance by excluding the top 5% hardest instances, while maintaining or increasing training efficacy. Empirical results across SR models and benchmarks show that descending sampling on pruned data can match or exceed full-dataset training, and that loss-based pruning outperforms Sobel-filter approaches. The approach offers meaningful reductions in data and compute requirements with little to no sacrifice in SR quality, and points to adaptive, uncertainty-aware core-sets as a promising direction for future work.

Abstract

In image Super-Resolution (SR), relying on large datasets for training is a double-edged sword. While offering rich training material, they also demand substantial computational and storage resources. In this work, we analyze dataset pruning to solve these challenges. We introduce a novel approach that reduces a dataset to a core-set of training samples, selected based on their loss values as determined by a simple pre-trained SR model. By focusing the training on just 50\% of the original dataset, specifically on the samples characterized by the highest loss values, we achieve results comparable to or surpassing those obtained from training on the entire dataset. Interestingly, our analysis reveals that the top 5\% of samples with the highest loss values negatively affect the training process. Excluding these samples and adjusting the selection to favor easier samples further enhances training outcomes. Our work opens new perspectives to the untapped potential of dataset pruning in image SR. It suggests that careful selection of training data based on loss-value metrics can lead to better SR models, challenging the conventional wisdom that more data inevitably leads to better performance.

A Study in Dataset Pruning for Image Super-Resolution

TL;DR

This paper tackles the data inefficiency of image SR by proposing a loss-value–based core-set pruning strategy. Using a simple pre-trained SRCNN to estimate reconstruction loss, it forms a 50% core-set by selecting high-loss samples and further improves performance by excluding the top 5% hardest instances, while maintaining or increasing training efficacy. Empirical results across SR models and benchmarks show that descending sampling on pruned data can match or exceed full-dataset training, and that loss-based pruning outperforms Sobel-filter approaches. The approach offers meaningful reductions in data and compute requirements with little to no sacrifice in SR quality, and points to adaptive, uncertainty-aware core-sets as a promising direction for future work.

Abstract

In image Super-Resolution (SR), relying on large datasets for training is a double-edged sword. While offering rich training material, they also demand substantial computational and storage resources. In this work, we analyze dataset pruning to solve these challenges. We introduce a novel approach that reduces a dataset to a core-set of training samples, selected based on their loss values as determined by a simple pre-trained SR model. By focusing the training on just 50\% of the original dataset, specifically on the samples characterized by the highest loss values, we achieve results comparable to or surpassing those obtained from training on the entire dataset. Interestingly, our analysis reveals that the top 5\% of samples with the highest loss values negatively affect the training process. Excluding these samples and adjusting the selection to favor easier samples further enhances training outcomes. Our work opens new perspectives to the untapped potential of dataset pruning in image SR. It suggests that careful selection of training data based on loss-value metrics can lead to better SR models, challenging the conventional wisdom that more data inevitably leads to better performance.
Paper Structure (13 sections, 3 equations, 4 figures, 3 tables)

This paper contains 13 sections, 3 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Core-Sets
  5. Loss-Value-based Sampling
  6. Experiments
  7. Datasets
  8. Results
  9. Ascending versus Descending Sampling.
  10. Loss-based versus Sobel-based
  11. Analyzing Loss Values.
  12. Benchmark with State-of-the-Art.
  13. Conclusion and Future Work

Figures (4)

  • Figure 1: Illustration of our loss-value-based core-set selection for image SR. Initially, the full dataset undergoes evaluation through a pre-trained SR model to calculate loss values for each image pair. These loss values are then sorted to identify samples with varying degrees of reconstruction difficulty. A pre-defined proportion $r$ of these samples is selected to form a core-set.
  • Figure 2: Comparison of top-selected samples by ascending and descending sampling. We can observe that descending sampling selects primarily training patches with high textural details, whereas ascending sampling focuses on monochromatic samples.
  • Figure 3: Cumulative Loss-Value Distribution (sorted). Vertical lines represent different descending sampling endpoints for 25%, 50%, and 75% sampling. The right side of the respective vertical line shows the loss values included and found within the corresponding core-sets.
  • Figure 4: Refined Core Set Proposal. This strategy selects the top 50% most challenging samples but modifies the selection by shifting the inclusion threshold by 5 % towards samples with lower loss values, aiming for a more balanced core-set. In other words, we keep 50 % of the hardest samples in our core-set after excluding the top 5 % from the dataset.