ReasonText

🌐 Project Page · 📄 arXiv (coming soon) · 💻 Code

ReasonText is the first benchmark for text-aware image super-resolution (TAISR) with the following two characteristics:

1. High quality — all annotations are produced by humans, with no automatic pipeline.
2. Per-instance difficulty — each text instance is labeled by difficulty: Level 1 (perception) and Level 2 (reasoning-required).

Contents

HR/                     # 513 high-resolution images (512×512 PNG)
LR/                     # 513 low-resolution images (128×128 PNG, ×4)
ReasonText-meta.json    # word-level annotations, keyed by image stem

ReasonText-meta.json maps each image stem (e.g. DRealSR__Canon_10__256_01) to:

{
  "hr_image": "HR/<stem>.png",
  "lr_image": "LR/<stem>.png",
  "annotations": [
    {
      "hr_coords":       [[x1,y1], ...],   // word-level polygon in HR pixel coordinates
      "text":            "BEST",           // ground-truth transcription
      "text_rotation":   0,                // 0 / 90 / 180 / 270 (counter-clockwise)
      "text_flipped":    false,
      "text_difficulty": "Level 1",        // Level 1 / Level 2 / Level 3
      "reason_types":    ["Context","Prior"] // Level 2 only; see "Reasoning types" below
    }
  ]
}

Difficulty levels

Each text instance is labeled by how it can be read:

Level	Type	Readable from	Instances
Level 1	perception	the cropped LR text region	1,754
Level 2	reasoning-required	the full LR image (via context, prior, or logic)	709

Text that human annotators could not identify even from the full LR image is labeled as Level 3 (total 1,522 instances), which is not the main focus of this benchmark.

Reasoning types (Level 2)

For a Level 2 instance, reason_types records which cue makes the text recoverable:

• Context — image context
• Prior — prior knowledge
• Logic — logical reasoning

Different annotators may recover the same text through different cues, so reason_types is not unique: it is the set of cues reported across annotators for that instance (a subset of the three).

Citation

If you find our work useful for your research, please consider citing our paper:

<fill in BibTeX>

License

ReasonText is released under CC BY-NC 4.0 (non-commercial research use only).