jiangbingo/textile-color-difference-dataset

Textile Color Difference Dataset

Dataset Description

Personal academic dataset for textile color difference research, containing standardized flatbed scans of fabric samples with CIE L*a*b* color statistics and CIE76 ΔE*ab computed against reference samples.

Data Source: Personal experimental collection.

Anonymization: All fabric style codes, internal directory names, and reference identifiers have been replaced with neutral IDs (fabric_NNN, batch_MMDD, ref_NN). No company, customer, factory, or business references remain in file names, fields, or metadata.

Dataset Structure

textile-color-difference-dataset/
├── annotations/                             # 123 JSON annotations (one per sample)
├── images/
│   ├── reference/                           # 13 reference standard samples (ref_01–ref_13)
│   └── batch/                               # 123 batch scans in 13 groups, 32 fabric types
├── metadata/
│   └── batch_summary.csv                    # 123-row summary matching annotations
├── scripts/
│   ├── build_hf_dataset.py                  # Regenerate metadata.jsonl / parquet
│   └── recalculate_delta_e.py               # Recalculate ΔE vs reference
├── metadata.jsonl                           # ImageFolder-compatible metadata
└── textile-color-difference-dataset.parquet # Tabular index (paths only, no embedded bytes)

Data Specifications

Field	Value
Scan Device	Epson V850 Pro
Light Source	D65 standard illuminant
Annotated Samples	123
Reference Samples	13
Batch Groups	13 scan sessions
Fabric Types	32
Image Format	JPG

Annotation Format

{
  "image_id": "2.jpg",
  "image_path": "images/batch/batch_0103/fabric_022/2.jpg",
  "metadata": {
    "fabric_type": "fabric_022",
    "scan_device": "Epson V850 Pro",
    "light_source": "D65",
    "batch_group": "batch_0103"
  },
  "color_statistics": {
    "L_mean": 31.76,
    "a_mean": 6.4,
    "b_mean": 3.25,
    "delta_e_mean": 2.02,
    "total_pixels": 1280987,
    "delta_e_reference": "ref_08.jpg",
    "delta_e_method": "CIE76_vs_reference"
  }
}

Fields

Field	Description
image_id	Image filename
image_path	Relative path to image (verified)
metadata.fabric_type	Anonymized fabric identifier (fabric_NNN)
metadata.batch_group	Scan session (batch_MMDD)
color_statistics.L_mean	Mean L* (lightness)
color_statistics.a_mean	Mean a* (green-red)
color_statistics.b_mean	Mean b* (blue-yellow)
color_statistics.delta_e_mean	CIE76 ΔE*ab vs reference sample
color_statistics.delta_e_reference	Reference image used for ΔE calculation
color_statistics.delta_e_method	Calculation method

ΔE Distribution

ΔE Range	Count	Interpretation
0-1	12	Imperceptible difference
1-3	32	Slight difference (typical textile pass threshold)
3-5	26	Noticeable difference
5-10	16	Significant difference
10-20	37	Large difference

Range: 0.23 — 16.81, Mean: 6.60

Label Status

Original pass/fail records were not preserved. Pseudo-labels can be derived from ΔE using a threshold (e.g. ΔE < 3 for textile pass → 44 pass / 79 fail).

Reference Mapping

13 reference images cover 32 fabric types through family matching. The full mapping is defined in scripts/recalculate_delta_e.py (FABRIC_TO_REF).

Quick Start

Load from Hugging Face

from datasets import load_dataset

ds = load_dataset("jiangbingo/textile-color-difference-dataset", split="train")
print(f"Samples: {len(ds)}")  # 123

sample = ds[0]
sample["image"]          # PIL Image
sample["fabric_type"]    # "fabric_001"
sample["delta_e_mean"]   # 2.02

Load from local directory

from datasets import load_dataset

ds = load_dataset("imagefolder", data_dir="./textile-color-difference-dataset", split="train")

Load tabular index from Parquet

The parquet file is a metadata index with repo-relative image paths (no embedded bytes):

from datasets import Dataset

ds = Dataset.from_parquet("textile-color-difference-dataset.parquet")

Work with annotations directly

import json
from PIL import Image

ann = json.load(open("annotations/batch_0103_fabric_022_2.json"))
img = Image.open(ann["image_path"])
cs = ann["color_statistics"]
print(f"Fabric: {ann['metadata']['fabric_type']}")
print(f"L*a*b*: ({cs['L_mean']}, {cs['a_mean']}, {cs['b_mean']})")
print(f"ΔE*ab: {cs['delta_e_mean']} (ref: {cs['delta_e_reference']})")

Compliance

• Fully anonymized: style codes, internal directory names, and identifiers replaced with neutral IDs
• No company, customer, factory, or business data
• Personal academic dataset
• MIT License

Limitations

• No human-verified pass/fail labels — pseudo-labels must be derived from ΔE thresholds
• 123 samples from a single scan device
• Some reference mappings are approximate (sub-variants share a parent reference)