Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TextureCrop: Enhancing Synthetic Image Detection through Texture-based Cropping

Despina Konstantinidou, Christos Koutlis, Symeon Papadopoulos

TL;DR

TextureCrop addresses the challenge of detecting synthetic content in high-resolution images by introducing a texture-driven pre-processing step that selects a fixed number of high-frequency crops via a sliding window and aggregates their detector scores. By prioritizing texture-rich regions using Global Histogram Entropy and averaging per-crop logits, TextureCrop consistently improves SID performance across multiple detectors and high-resolution datasets, while reducing memory and compute relative to full-image processing. The method yields meaningful gains in BA, AP, and AUC (e.g., up to 12.1% BA, 18.2% AP, and 14.9% AUC over resizing) and demonstrates robustness across datasets like Forensynths, Synthbuster, and TWIGMA, with notable efficiency improvements on large images. Its simplicity and compatibility with existing SID models position TextureCrop as a practical, plug-in enhancement for real-world synthetic image mitigation and surveillance workflows.

Abstract

Generative AI technologies produce increasingly realistic imagery, which, despite its potential for creative applications, can also be misused to produce misleading and harmful content. This renders Synthetic Image Detection (SID) methods essential for identifying AI-generated content online. State-of-the-art SID methods typically resize or center-crop input images due to architectural or computational constraints, which hampers the detection of artifacts that appear in high-resolution images. To address this limitation, we propose TextureCrop, an image pre-processing component that can be plugged in any pre-trained SID model to improve its performance. By focusing on high-frequency image parts where generative artifacts are prevalent, TextureCrop enhances SID performance with manageable memory requirements. Experimental results demonstrate a consistent improvement in AUC across various detectors by 6.1% compared to center cropping and by 15% compared to resizing, across high-resolution images from the Forensynths, Synthbuster and TWIGMA datasets. Code available at https : //github.com/mever-team/texture-crop.

TextureCrop: Enhancing Synthetic Image Detection through Texture-based Cropping

TL;DR

TextureCrop addresses the challenge of detecting synthetic content in high-resolution images by introducing a texture-driven pre-processing step that selects a fixed number of high-frequency crops via a sliding window and aggregates their detector scores. By prioritizing texture-rich regions using Global Histogram Entropy and averaging per-crop logits, TextureCrop consistently improves SID performance across multiple detectors and high-resolution datasets, while reducing memory and compute relative to full-image processing. The method yields meaningful gains in BA, AP, and AUC (e.g., up to 12.1% BA, 18.2% AP, and 14.9% AUC over resizing) and demonstrates robustness across datasets like Forensynths, Synthbuster, and TWIGMA, with notable efficiency improvements on large images. Its simplicity and compatibility with existing SID models position TextureCrop as a practical, plug-in enhancement for real-world synthetic image mitigation and surveillance workflows.

Abstract

Generative AI technologies produce increasingly realistic imagery, which, despite its potential for creative applications, can also be misused to produce misleading and harmful content. This renders Synthetic Image Detection (SID) methods essential for identifying AI-generated content online. State-of-the-art SID methods typically resize or center-crop input images due to architectural or computational constraints, which hampers the detection of artifacts that appear in high-resolution images. To address this limitation, we propose TextureCrop, an image pre-processing component that can be plugged in any pre-trained SID model to improve its performance. By focusing on high-frequency image parts where generative artifacts are prevalent, TextureCrop enhances SID performance with manageable memory requirements. Experimental results demonstrate a consistent improvement in AUC across various detectors by 6.1% compared to center cropping and by 15% compared to resizing, across high-resolution images from the Forensynths, Synthbuster and TWIGMA datasets. Code available at https : //github.com/mever-team/texture-crop.
Paper Structure (30 sections, 6 equations, 7 figures, 2 tables)

This paper contains 30 sections, 6 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Overview
  3. Related Work
  4. Synthetic image generation
  5. Synthetic image detection
  6. Proposed Approach
  7. Problem Formulation
  8. Selecting texture-rich image crops
  9. Parameters
  10. Stride ($\sigma$)
  11. Window Size ($\omega$)
  12. Number of Crops ($\nu$)
  13. Texture Selection Measure ($\mu$)
  14. Aggregation method
  15. Experimental Setup
  16. ...and 15 more sections

Figures (7)

  • Figure 1: Example of the crops created from a 1792x1024 image by using (a) resizing, (b) center cropping, (c) ten cropping and (d) texture cropping. Images (a) and (b) have the same dimensions ($224\times224$) as the 10 crops shown in (c) and (d).
  • Figure 2: Overview of the TextureCrop Pipeline.
  • Figure 3: Box plot of (a) BA, (b) AP and (c) AUC distribution across detection methods for different values of the stride parameter $\sigma$, and (d) BA, (e) AP and (f) AUC distribution for different values of the number of crops parameter $\nu$.
  • Figure 4: Example of the crops created from a 1792x1024 image (a) using texture cropping with global histogram entropy (b), local entropy (c), standard deviation (d), inverse autocorrelation (e) and texture diversity (f) as a metric.
  • Figure 5: Performance (AUC) of TextureCrop for different texture selection criteria $\mu$ across different detectors.
  • ...and 2 more figures