OCR in the Wild: Cloud and Edge Implementations

Overview

This repository demonstrates implementations of Optical Character Recognition (OCR) for text in the wild, designed for both cloud and edge deployment scenarios. Our goal is to provide flexible, efficient OCR solutions that can be adapted to various use cases and hardware constraints.

OCR Process

Our OCR system employs a two-step process for accurate text recognition in natural scenes:

1. Text Detection: This step identifies and locates all instances of text within an image. It produces bounding boxes around detected text areas.

2. Text Recognition: After detection, the identified text regions are cropped and batched. These batches are then passed through a recognition model that converts the image of text into machine-readable characters.

Repository Structure

This repository is divided into two main sections:

1. Cloud OCR: A scalable, high-performance OCR system designed for cloud deployment.
2. Edge OCR: A lightweight, efficient OCR model suitable for deployment on edge devices.

Cloud OCR Implementation

Our cloud-based OCR system leverages pre-trained models from PaddlePaddleOCR, converted to ONNX format for improved performance and compatibility. It uses Triton Inference Server for efficient model serving and exposes its functionality through a Flask API.

Key features of the cloud implementation:

• Uses state-of-the-art PaddlePaddleOCR models for text detection and recognition
• ONNX-format models for cross-platform compatibility
• Triton Inference Server for scalable model serving
• Flask API for easy integration with various applications
• Designed for deployment in cloud environments
• Efficient batching of detected text regions for recognition

For more details, see the Cloud OCR README.

Edge OCR Implementation

Our edge-based OCR system is designed to run efficiently on resource-constrained devices. It uses custom, lightweight model architectures for both detection and recognition that prioritize speed and low memory footprint without significantly compromising accuracy.

Key features of the edge implementation:

• Custom CNN architectures optimized for edge devices
• Uses only convolutional operations for wide hardware compatibility
• Text detection model designed for quick region proposals
• Text recognition model trained with CTC loss for efficient sequence learning
• Configurable for different input sizes and character sets
• Easily convertible to ONNX and TensorRT for further optimization
• Optimized pipeline for efficient processing of detected text regions

For more details, see the Edge OCR README.

Comparison

Feature	Cloud OCR	Edge OCR
Deployment Target	High-performance cloud servers	Resource-constrained edge devices
Model Source	Pre-trained PaddlePaddleOCR	Custom-trained lightweight models (Compatible ops)
Text Detection	Advanced detection model	Yolov8
Text Recognition	Advanced LSTM-based model	Compact CNN with CTC loss
Inference Server	Triton Inference Server	Direct inference
Primary Advantages	High accuracy, scalability, centralized	Low latency, minimal resource usage, privacy, offline, cost

Getting Started

To get started with either the cloud or edge implementation:

1. Clone this repository:

   git clone https://github.com/your-username/ocr-implementations.git
   

2. Navigate to either the cloud or edge directory:

   cd ocr-implementations/cloud
   # or
   cd ocr-implementations/edge
   

3. Follow the setup instructions in the respective README files.