sabaridsnfuji/rvl-cdip-filtered

RVL-CDIP Filtered Dataset

This dataset contains filtered images from the RVL-CDIP dataset, focusing on 4 specific document types.

Dataset Summary

A filtered subset of the RVL-CDIP (Ryerson Vision Lab Complex Document Information Processing) dataset containing 100,000 images across 4 document categories. Each image is stored as base64-encoded data in Parquet format for efficient processing.

Classes

Label	Class Name	Description
0	letter	Personal and business letters
1	form	Forms and applications
2	email	Email messages
14	resume	Resumes and CVs

Dataset Statistics

Train Split

• 0_letter: 20,106 images
• 1_form: 19,957 images
• 2_email: 19,954 images
• 14_resume: 20,037 images
• Total train: 80,054 images

Validation Split

• 0_letter: 2,430 images
• 1_form: 2,537 images
• 2_email: 2,530 images
• 14_resume: 2,426 images
• Total validation: 9,923 images

Test Split

• 0_letter: 2,464 images
• 1_form: 2,506 images
• 2_email: 2,516 images
• 14_resume: 2,537 images
• Total test: 10,023 images

Grand Total: 100,000 images

Dataset Structure

The dataset is provided in Parquet format with the following columns:

• image_path: Original path of the image file
• image_base64: Base64-encoded image data
• label: Integer label (0, 1, 2, or 14)
• class_name: Human-readable class name
• width: Image width in pixels
• height: Image height in pixels
• format: Image format (jpg, png, etc.)
• file_size_bytes: Original file size in bytes
• split: Dataset split (train/validation/test)

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset
import base64
from PIL import Image
import io

# Load the dataset
dataset = load_dataset("sabaridsnfuji/rvl-cdip-filtered")

# Access splits
train_data = dataset['train']
val_data = dataset['validation'] 
test_data = dataset['test']

# Decode an image
def decode_image(base64_string):
    image_data = base64.b64decode(base64_string)
    image = Image.open(io.BytesIO(image_data))
    return image

# Example: Get first image from training set
first_record = train_data[0]
image = decode_image(first_record['image_base64'])
label = first_record['label']
class_name = first_record['class_name']

Loading with Pandas

import pandas as pd
import base64
from PIL import Image
import io

# Load specific split
df = pd.read_parquet('train.parquet')

# Decode images
def decode_image(base64_string):
    image_data = base64.b64decode(base64_string)
    return Image.open(io.BytesIO(image_data))

# Apply to get images
df['image'] = df['image_base64'].apply(decode_image)

File Formats

• train.parquet: Training set
• validation.parquet: Validation set
• test.parquet: Test set

Each Parquet file contains the complete image data and metadata, making the dataset self-contained and easy to work with.

Original Dataset

This is a filtered subset of the RVL-CDIP dataset:

• Original paper: "RVL-CDIP: A New Dataset for Cross-Domain Document Layout Analysis"
• Original dataset: Contains 16 classes, we selected 4 most relevant ones
• Maintains: Original train/validation/test splits
• Citation:

  @inproceedings{{harley2015evaluation,
    title={{Evaluation of deep convolutional nets for document image classification and retrieval}},
    author={{Harley, Adam W and Ufkes, Alex and Derpanis, Konstantinos G}},
    booktitle={{International Conference on Document Analysis and Recognition}},
    pages={{991--995}},
    year={{2015}},
    organization={{IEEE}}
  }}
  

License

This dataset follows the same license terms as the original RVL-CDIP dataset.

Dataset Creation

The dataset was created by:

1. Filtering the original RVL-CDIP dataset for specific document types
2. Converting images to base64 encoding for efficient storage
3. Adding metadata (dimensions, format, file size)
4. Saving in Parquet format for optimal performance

Use Cases

• Document classification
• Computer vision research
• Transfer learning for document analysis
• Multi-class classification benchmarking
• Document understanding systems