Build Your Own OCR Engine for Wingdings

Optical Character Recognition (OCR) has revolutionized the best way we work together with textual knowledge in actual life, enabling machines to learn and interpret textual content from photographs, scanned paperwork, and handwritten notes. From digitizing books and automating knowledge entry to real-time textual content translation in augmented actuality, OCR purposes are extremely various and impactful. A few of its utility might embody:

• Doc Digitization: Converts bodily paperwork into editable and searchable digital codecs.
• Bill Scanning: Extracts particulars like quantities, dates, and vendor names for automated processing.
• Knowledge Entry Automation: Quickens workflows by extracting textual content from varieties and receipts.
• Actual-Time Translation: Interprets international textual content from photographs or video streams in augmented actuality.
• License Plate Recognition: Identifies automobiles in visitors techniques and parking administration.
• Accessibility Instruments: Converts textual content to speech for visually impaired people.
• Archiving and Preservation: Digitizes historic paperwork for storage and analysis.

On this put up, we take OCR a step additional by constructing a customized OCR mannequin for recognizing textual content within the Wingdings font—a symbolic font with distinctive characters usually utilized in artistic and technical contexts. Whereas conventional OCR fashions are skilled for traditional textual content, this practice mannequin bridges the hole for area of interest purposes, unlocking prospects for translating symbolic textual content into readable English, whether or not for accessibility, design, or archival functions. By means of this, we display the facility of OCR to adapt and cater to specialised use instances within the fashionable world.

For builders and managers seeking to streamline doc workflows reminiscent of OCR extraction and past, instruments just like the Nanonets PDF AI provide invaluable integration choices. Coupled with state-of-the-art LLM capabilities, these can considerably improve your workflows, guaranteeing environment friendly knowledge dealing with. Moreover, instruments like Nanonets’ PDF Summarizer can additional automate processes by summarizing prolonged paperwork.

Is There a Want for Customized OCR within the Age of Imaginative and prescient-Language Fashions?

Imaginative and prescient-language fashions, reminiscent of Flamingo, Qwen2-VL, have revolutionized how machines perceive photographs and textual content by bridging the hole between the 2 modalities. They will course of and motive about photographs and related textual content in a extra generalized method.

Regardless of their spectacular capabilities, there stays a necessity for customized OCR techniques in particular situations, primarily attributable to:

• Accuracy for Particular Languages or Scripts: Many vision-language fashions give attention to widely-used languages. Customized OCR can deal with low-resource or regional languages, together with Indic scripts, calligraphy, or underrepresented dialects.
• Light-weight and Useful resource-Constrained Environments: Customized OCR fashions might be optimized for edge gadgets with restricted computational energy, reminiscent of embedded techniques or cell purposes. Imaginative and prescient-language fashions, in distinction, are sometimes too resource-intensive for such use instances. For real-time or high-volume purposes, reminiscent of bill processing or automated doc evaluation, customized OCR options might be tailor-made for velocity and accuracy.
• Knowledge Privateness and Safety: Sure industries, reminiscent of healthcare or finance, require OCR options that function offline or inside personal infrastructures to fulfill strict knowledge privateness laws. Customized OCR ensures compliance, whereas cloud-based vision-language fashions may introduce safety issues.
• Value-Effectiveness: Deploying and fine-tuning large vision-language fashions might be cost-prohibitive for small-scale companies or particular tasks. Customized OCR generally is a extra inexpensive and centered different.

Construct a Customized OCR Mannequin for Wingdings

To discover the potential of customized OCR techniques, we’ll construct an OCR engine particularly for the Wingdings font.

Beneath are the steps and parts we’ll comply with:

• Generate a customized dataset of Wingdings font photographs paired with their corresponding labels in English phrases.
• Create a customized OCR mannequin able to recognizing symbols within the Wingdings font. We’ll use the Imaginative and prescient Transformer for Scene Textual content Recognition (ViTSTR), a state-of-the-art structure designed for image-captioning duties. Not like conventional CNN-based fashions, ViTSTR leverages the transformer structure, which excels at capturing long-range dependencies in photographs, making it excellent for recognizing advanced textual content buildings, together with the intricate patterns of Wingdings fonts.
• Prepare the mannequin on the customized dataset of Wingdings symbols.
• Take a look at the mannequin on unseen knowledge to judge its accuracy.

For this undertaking, we’ll make the most of Google Colab for coaching, leveraging its 16 GB T4 GPU for quicker computation.

Making a Wingdings Dataset

What’s Wingdings?

Wingdings is a symbolic font developed by Microsoft that consists of a group of icons, shapes, and pictograms as a substitute of conventional alphanumeric characters. Launched in 1990, Wingdings maps keyboard inputs to graphical symbols, reminiscent of arrows, smiley faces, checkmarks, and different ornamental icons. It’s usually used for design functions, visible communication, or as a playful font in digital content material.

Resulting from its symbolic nature, decoding Wingdings textual content programmatically poses a problem, making it an fascinating use case for customized OCR techniques.

Dataset Creation

Since no present dataset is out there for Optical Character Recognition (OCR) in Wingdings font, we created one from scratch. The method includes producing photographs of phrases within the Wingdings font and mapping them to their corresponding English phrases.

To realize this, we used the Wingdings Translator to transform English phrases into their Wingdings representations. For every transformed phrase, a picture was manually generated and saved in a folder named “wingdings_word_images”.

Moreover, we create a “metadata.csv” file to take care of a structured report of the dataset together with the picture path. This file accommodates two columns:

1. Picture Path: Specifies the file path for every picture within the dataset.
2. English Phrase: Lists the corresponding English phrase for every Wingdings illustration.

The dataset might be downloaded from this link.

Preprocessing the Dataset

The photographs within the dataset range in measurement because of the handbook creation course of. To make sure uniformity and compatibility with OCR fashions, we preprocess the pictures by resizing and padding them.

import pandas as pd
import numpy as np
from PIL import Picture
import os
from tqdm import tqdm

def pad_image(picture, target_size=(224, 224)):
    """Pad picture to focus on measurement whereas sustaining side ratio"""
    if picture.mode != 'RGB':
        picture = picture.convert('RGB')
    
    # Get present measurement
    width, peak = picture.measurement
    
    # Calculate padding
    aspect_ratio = width / peak
    if aspect_ratio > 1:
        # Width is bigger
        new_width = target_size[0]
        new_height = int(new_width / aspect_ratio)
    else:
        # Peak is bigger
        new_height = target_size[1]
        new_width = int(new_height * aspect_ratio)
    
    # Resize picture sustaining side ratio
    picture = picture.resize((new_width, new_height), Picture.Resampling.LANCZOS)
    
    # Create new picture with padding
    new_image = Picture.new('RGB', target_size, (255, 255, 255))
    
    # Paste resized picture in heart
    paste_x = (target_size[0] - new_width) // 2
    paste_y = (target_size[1] - new_height) // 2
    new_image.paste(picture, (paste_x, paste_y))
    
    return new_image

# Learn the metadata
df = pd.read_csv('metadata.csv')

# Create output listing for processed photographs
processed_dir="processed_images"
os.makedirs(processed_dir, exist_ok=True)

# Course of every picture
new_paths = []
for idx, row in tqdm(df.iterrows(), whole=len(df), desc="Processing photographs"):
    # Load picture
    img_path = row['image_path']
    img = Picture.open(img_path)
    
    # Pad picture
    processed_img = pad_image(img)
    
    # Save processed picture
    new_path = os.path.be a part of(processed_dir, f'processed_{os.path.basename(img_path)}')
    processed_img.save(new_path)
    new_paths.append(new_path)

# Replace dataframe with new paths
df['processed_image_path'] = new_paths
df.to_csv('processed_metadata.csv', index=False)

print("Picture preprocessing accomplished!")
print(f"Complete photographs processed: {len(df)}")

First, every picture is resized to a set peak whereas sustaining its side ratio to protect the visible construction of the Wingdings characters. Subsequent, we apply padding to make all photographs the identical dimensions, sometimes a sq. form, to suit the enter necessities of neural networks. The padding is added symmetrically across the resized picture, with the background shade matching the unique picture’s background.

Splitting the Dataset

The dataset is split into three subsets: coaching (70%), validation (dev) (15%), and testing (15%). The coaching set is used to show the mannequin, the validation set helps fine-tune hyperparameters and monitor overfitting, and the check set evaluates the mannequin’s efficiency on unseen knowledge. This random break up ensures every subset is various and consultant, selling efficient generalization.

import pandas as pd
from sklearn.model_selection import train_test_split

# Learn the processed metadata
df = pd.read_csv('processed_metadata.csv')

# First break up: prepare and momentary
train_df, temp_df = train_test_split(df, train_size=0.7, random_state=42)

# Second break up: validation and check from momentary
val_df, test_df = train_test_split(temp_df, train_size=0.5, random_state=42)

# Save splits to CSV
train_df.to_csv('prepare.csv', index=False)
val_df.to_csv('val.csv', index=False)
test_df.to_csv('check.csv', index=False)

print("Knowledge break up statistics:")
print(f"Coaching samples: {len(train_df)}")
print(f"Validation samples: {len(val_df)}")
print(f"Take a look at samples: {len(test_df)}")

Visualizing the Dataset

To raised perceive the dataset, we visualize samples from every break up. Particularly, we show 5 examples from the coaching set, 5 from the validation set, and 5 from the check set. Every visualization consists of the Wingdings textual content as a picture alongside its corresponding label in English. This step gives a transparent overview of the information distribution throughout the splits and ensures the correctness of the dataset mappings.

import matplotlib.pyplot as plt
from PIL import Picture
import pandas as pd

def plot_samples(df, num_samples=5, title="Pattern Photos"):
    # Set bigger font sizes
    plt.rcParams.replace({
        'font.measurement': 14,          # Base font measurement
        'axes.titlesize': 16,     # Subplot title font measurement
        'determine.titlesize': 20    # Predominant title font measurement
    })
    
    fig, axes = plt.subplots(1, num_samples, figsize=(20, 4))
    fig.suptitle(title, fontsize=20, y=1.05)
    
    # Randomly pattern photographs
    sample_df = df.pattern(n=num_samples)
    
    for idx, (_, row) in enumerate(sample_df.iterrows()):
        img = Picture.open(row['processed_image_path'])
        axes[idx].imshow(img)
        axes[idx].set_title(f"Label: {row['english_word_label']}", fontsize=16, pad=10)
        axes[idx].axis('off')
    
    plt.tight_layout()
    plt.present()

# Load splits
train_df = pd.read_csv('prepare.csv')
val_df = pd.read_csv('val.csv')
test_df = pd.read_csv('check.csv')

# Plot samples from every break up
plot_samples(train_df, title="Coaching Samples")
plot_samples(val_df, title="Validation Samples")
plot_samples(test_df, title="Take a look at Samples")

Samples from the information are visualised as:

Prepare an OCR Mannequin

First we have to import the required libraries and dependencies:

import torch
import torch.nn as nn
from torch.utils.knowledge import Dataset, DataLoader
from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
from PIL import Picture
import pandas as pd
from tqdm import tqdm

Mannequin Coaching with ViTSTR

We use a Imaginative and prescient Encoder-Decoder mannequin, particularly ViTSTR (Imaginative and prescient Transformer for Scene Textual content Recognition). We fine-tune it for our Wingdings OCR activity. The encoder processes the Wingdings textual content photographs utilizing a ViT (Imaginative and prescient Transformer) spine, whereas the decoder generates the corresponding English phrase labels.

Throughout coaching, the mannequin learns to map pixel-level data from the pictures to significant English textual content. The coaching and validation losses are monitored to evaluate mannequin efficiency, guaranteeing it generalizes properly. After coaching, the fine-tuned mannequin is saved for inference on unseen Wingdings textual content photographs. We use pre-trained parts from Hugging Face for our OCR pipeline and effective tune them. The ViTImageProcessor prepares photographs for the Imaginative and prescient Transformer (ViT) encoder, whereas the bert-base-uncased tokenizer processes English textual content labels for the decoder. The VisionEncoderDecoderModel, combining a ViT encoder and GPT-2 decoder, is fine-tuned for picture captioning duties, making it excellent for studying the Wingdings-to-English mapping.

class WingdingsDataset(Dataset):
    def __init__(self, csv_path, processor, tokenizer):
        self.df = pd.read_csv(csv_path)
        self.processor = processor
        self.tokenizer = tokenizer
    
    def __len__(self):
        return len(self.df)
    
    def __getitem__(self, idx):
        row = self.df.iloc[idx]
        picture = Picture.open(row['processed_image_path'])
        label = row['english_word_label']
        
        # Course of picture
        pixel_values = self.processor(picture, return_tensors="pt").pixel_values
        
        # Course of label
        encoding = self.tokenizer(
            label,
            padding="max_length",
            max_length=16,
            truncation=True,
            return_tensors="pt"
        )
        
        return {
            'pixel_values': pixel_values.squeeze(),
            'labels': encoding.input_ids.squeeze(),
            'textual content': label
        }

def train_epoch(mannequin, dataloader, optimizer, system):
    mannequin.prepare()
    total_loss = 0
    progress_bar = tqdm(dataloader, desc="Coaching")
    
    for batch in progress_bar:
        pixel_values = batch['pixel_values'].to(system)
        labels = batch['labels'].to(system)
        
        outputs = mannequin(pixel_values=pixel_values, labels=labels)
        loss = outputs.loss
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        total_loss += loss.merchandise()
        progress_bar.set_postfix({'loss': loss.merchandise()})
    
    return total_loss / len(dataloader)

def validate(mannequin, dataloader, system):
    mannequin.eval()
    total_loss = 0
    
    with torch.no_grad():
        for batch in tqdm(dataloader, desc="Validating"):
            pixel_values = batch['pixel_values'].to(system)
            labels = batch['labels'].to(system)
            
            outputs = mannequin(pixel_values=pixel_values, labels=labels)
            loss = outputs.loss
            total_loss += loss.merchandise()
    
    return total_loss / len(dataloader)

# Initialize fashions and tokenizers
processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
mannequin = VisionEncoderDecoderModel.from_pretrained("nlpconnect/vit-gpt2-image-captioning")

# Create datasets
train_dataset = WingdingsDataset('prepare.csv', processor, tokenizer)
val_dataset = WingdingsDataset('val.csv', processor, tokenizer)

# Create dataloaders
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=32)

# Setup coaching
system = torch.system('cuda' if torch.cuda.is_available() else 'cpu')
mannequin = mannequin.to(system)
optimizer = torch.optim.AdamW(mannequin.parameters(), lr=5e-5)
num_epochs = 20 #(change in accordance with want)

# Coaching loop
for epoch in vary(num_epochs):
    print(f"nEpoch {epoch+1}/{num_epochs}")
    
    train_loss = train_epoch(mannequin, train_loader, optimizer, system)
    val_loss = validate(mannequin, val_loader, system)
    
    print(f"Coaching Loss: {train_loss:.4f}")
    print(f"Validation Loss: {val_loss:.4f}")

# Save the mannequin
mannequin.save_pretrained('wingdings_ocr_model')
print("nTraining accomplished and mannequin saved!")

The coaching is carried for 20 epochs in Google Colab. Though it offers honest end result with 20 epochs, it is a hyper parameter and might be elevated to achieve higher accuracy. Dropout, Picture Augmentation and Batch Normalization are a couple of extra hyper-parameters one can play with to make sure mannequin is just not overfitting. The coaching stats and the loss and accuracy curve for prepare and validation units on first and final epochs are given under:

Epoch 1/20
Coaching: 100%|██████████| 22/22 [00:36<00:00,  1.64s/it, loss=1.13]
Validating: 100%|██████████| 5/5 [00:02<00:00,  1.71it/s]
Coaching Loss: 2.2776
Validation Loss: 1.0183

..........
..........
..........
..........

Epoch 20/20
Coaching: 100%|██████████| 22/22 [00:35<00:00,  1.61s/it, loss=0.0316]
Validating: 100%|██████████| 5/5 [00:02<00:00,  1.73it/s]
Coaching Loss: 0.0246
Validation Loss: 0.5970

Coaching accomplished and mannequin saved!

Utilizing the Saved Mannequin

As soon as the mannequin has been skilled and saved, you possibly can simply load it for inference on new Wingdings photographs. The check.csv file created throughout preprocessing is used to create the test_dataset. Right here’s the code to load the saved mannequin and make predictions:

# Load the skilled mannequin
mannequin = VisionEncoderDecoderModel.from_pretrained('wingdings_ocr_model')
processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224-in21k")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

# Create check dataset and dataloader
test_dataset = WingdingsDataset('check.csv', processor, tokenizer)
test_loader = DataLoader(test_dataset, batch_size=32)

Mannequin Analysis

After coaching, we consider the mannequin’s efficiency on the check break up to measure its efficiency. To achieve insights into the mannequin’s efficiency, we randomly choose 10 samples from the check break up. For every pattern, we show the true label (English phrase) alongside the mannequin’s prediction and verify in the event that they match.

import seaborn as sns
import matplotlib.pyplot as plt
from PIL import Picture

def plot_prediction_samples(image_paths, true_labels, pred_labels, num_samples=10):
    # Set determine measurement and font sizes
    plt.rcParams.replace({
        'font.measurement': 14,
        'axes.titlesize': 18,
        'determine.titlesize': 22
    })
    
    # Calculate grid dimensions
    num_rows = 2
    num_cols = 5
    num_samples = min(num_samples, len(image_paths))
    
    # Create determine
    fig, axes = plt.subplots(num_rows, num_cols, figsize=(20, 8))
    fig.suptitle('Pattern Predictions from Take a look at Set', fontsize=22, y=1.05)
    
    # Flatten axes for simpler indexing
    axes_flat = axes.flatten()
    
    for i in vary(num_samples):
        ax = axes_flat[i]
        
        # Load and show picture
        img = Picture.open(image_paths[i])
        ax.imshow(img)
        
        # Create label textual content
        true_text = f"True: {true_labels[i]}"
        pred_text = f"Pred: {pred_labels[i]}"
        
        # Set shade primarily based on correctness
        shade="inexperienced" if true_labels[i] == pred_labels[i] else 'crimson'
        
        # Add textual content above picture
        ax.set_title(f"{true_text}n{pred_text}", 
                    fontsize=14,
                    shade=shade,
                    pad=10,
                    bbox=dict(facecolor="white", 
                             alpha=0.8,
                             edgecolor="none",
                             pad=3))
        
        # Take away axes
        ax.axis('off')
    
    # Take away any empty subplots
    for i in vary(num_samples, num_rows * num_cols):
        fig.delaxes(axes_flat[i])
    
    plt.tight_layout()
    plt.present()
    
# Analysis
system = torch.system('cuda' if torch.cuda.is_available() else 'cpu')
mannequin = mannequin.to(system)
mannequin.eval()

predictions = []
ground_truth = []
image_paths = []

with torch.no_grad():
    for batch in tqdm(test_loader, desc="Evaluating"):
        pixel_values = batch['pixel_values'].to(system)
        texts = batch['text']
        
        outputs = mannequin.generate(pixel_values)
        pred_texts = tokenizer.batch_decode(outputs, skip_special_tokens=True)
        
        predictions.prolong(pred_texts)
        ground_truth.prolong(texts)
        image_paths.prolong([row['processed_image_path'] for _, row in test_dataset.df.iterrows()])

# Calculate and print accuracy
accuracy = accuracy_score(ground_truth, predictions)
print(f"nTest Accuracy: {accuracy:.4f}")

# Show pattern predictions in grid
print("nDisplaying pattern predictions:")
plot_prediction_samples(image_paths, ground_truth, predictions)    

The analysis offers the next output:

Analysing the output given by the mannequin, we discover that the predictions match the reference/authentic labels pretty properly. Though the final prediction is right it’s displayed in crimson due to the areas within the generated textual content.

All of the code and dataset used above might be discovered on this Github repository. And the top to finish coaching might be discovered within the following colab pocket book

Dialogue

Once we see the outputs, it turns into clear that the mannequin performs very well. The anticipated labels are correct, and the visible comparability with the true labels demonstrates the mannequin’s robust functionality in recognizing the right lessons.

The mannequin’s wonderful efficiency may very well be attributed to the strong structure of the Imaginative and prescient Transformer for Scene Textual content Recognition (ViTSTR). ViTSTR stands out attributable to its capability to seamlessly mix the facility of Imaginative and prescient Transformers (ViT) with language fashions for textual content recognition duties.

A comparability may very well be made by experimenting with completely different ViT structure sizes, reminiscent of various the variety of layers, embedding dimensions, or the variety of consideration heads. Fashions like ViT-Base, ViT-Giant, and ViT-Big might be examined, together with different architectures like:

• DeiT (Knowledge-efficient Picture Transformer)
• Swin Transformer

By evaluating these fashions of various scales, we are able to determine which structure is probably the most environment friendly by way of efficiency and computational assets. It will assist decide the optimum mannequin measurement that balances accuracy and effectivity for the given activity.

For duties like extracting data from paperwork, instruments reminiscent of Nanonets’ Chat with PDF have evaluated and used a number of state of the LLMs together with customized in-house skilled fashions and may provide a dependable technique to work together with content material, guaranteeing correct knowledge extraction with out threat of misrepresentation.