README.md

---
library_name: onnx
license: apache-2.0
tags:
  - foundation
  - amd
  - rocm
  - optical-character-recognition
pipeline_tag: image-to-text
---

![](https://huggingface.co/AMD-PAVS-AI/easyocr/resolve/main/easyocr.png)

# EasyOCR: Optimized for AMD ROCm

EasyOCR is a ready-to-use optical character recognition (OCR) pipeline combining a CRAFT text-detection model with a CRNN text-recognition model, supporting 80+ languages. This repository packages inference for OCR (text detection + recognition) using **ONNX Runtime**, exported and validated for **AMD ROCm** so it runs efficiently on AMD GPUs, CPUs, and NPUs.

This is based on the implementation of EasyOCR found [here](https://github.com/JaidedAI/EasyOCR).
This repository contains configurations and scripts optimized for **AMD® ROCm™** platforms. You can use the [easyocr AMD scripts](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/easyocr) to reproduce results or export with custom configurations. More details on model performance can be found [here](#accuracy-pipeline).

---

## Task Overview

**Task:** Optical character recognition (text detection + recognition)

**Dataset:** COCO-Text v2 subset (813 real-world images / 8,377 ground-truth text regions, sampled from COCO2014 images with ≥5 legible text regions each)

**Output metrics:** Word Accuracy, Character Similarity, Detection Precision/Recall/F1, FPS

> **Backend note:** Recognition always runs on CPU regardless of target device — its ~3ms latency is negligible next to detection, so the cross-device switch overhead isn't worth it — while detection runs on the target device's execution provider.

---

## AMD ROCm Optimization

This model export has been adapted and validated for **AMD Instinct™ / Radeon™ GPUs** running **ROCm**, as well as AMD CPUs and AMD Ryzen AI NPUs. Key points:

- Validated backends: **ONNX Runtime** across CPU, GPU (MIGraphX execution provider), and NPU (VitisAI execution provider, auto-quantized internally).
- Both detection (CRAFT) and recognition (CRNN) models are exported to ONNX with static shapes for cross-device compatibility.
- No code changes required versus the upstream EasyOCR implementation — only environment/runtime configuration differs.

| Runtime | Precision | Backend | Hardware | Notes |
|---|---|---|---|---|
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | CPU Execution Provider | AMD CPU | Recognition always runs here |
| ONNX Runtime | FP32 / FP16 / BF16 / INT8 | MIGraphX Execution Provider | AMD Instinct™ / Radeon™ GPU (ROCm) | Detection only |
| ONNX Runtime | Auto | VitisAI Execution Provider | AMD Ryzen AI NPU | Detection only; auto-quantized internally |

---

## Getting Started

For setup instructions, evaluation scripts, and custom configuration options, see the [easyocr on GitHub](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/easyocr).

---

## Model Details

**Model Type:** Optical character recognition pipeline (text detection + text recognition)

**Base Model:** CRAFT (detection) + CRNN (recognition), from [EasyOCR](https://github.com/JaidedAI/EasyOCR)

**Model Stats:**
- Detection input: `(1, 3, 640, 640)` float32; outputs region/link scores `(1, 320, 320, 2)` and feature map `(1, 32, 320, 320)`
- Recognition input: `(1, 1, 64, 200)` float32; output CTC logits `(1, 49, 97)`
- Precision tested: FP32, FP16, BF16, INT8 (CPU/GPU); auto-quantized (NPU)

---

## Accuracy Pipeline

Higher values mean predictions agree more closely with ground truth across all metrics below — 1.0 is perfect, 0.0 is no correct matches. Word Accuracy is intentionally low in absolute terms: COCO-Text's images are real-world "text in the wild" photos (street signs, product labels, graffiti) with small, rotated, and low-contrast text, unlike clean scanned documents.

### Metrics Explained

| Metric | Description |
|--------|-------------|
| Word Accuracy | Fraction of IoU-matched detections whose recognized text exactly matches (case-insensitive) the ground-truth string. The strictest text metric — a single wrong character fails the match. |
| Character Similarity | Mean character-level similarity ratio between recognized and ground-truth text across matched detections. More forgiving than Word Accuracy. |
| Detection Precision | Of all text regions the model detected, what fraction matched a real ground-truth region (IoU ≥ 0.5). |
| Detection Recall | Of all ground-truth text regions, what fraction the model detected. |
| Detection F1 | Harmonic mean of Detection Precision and Recall. |

### Accuracy Results

**Full Dataset Evaluation (COCO-Text v2, 813 images, 8,377 ground-truth text regions)** — filled from `runs/quality_<device>.json`; run `make metrics` to refresh:

<!-- accuracy-table-start -->
| Device | Precision | Word Accuracy | Char Similarity | Det. Precision | Det. Recall | Det. F1 |
|--------|-----------|---------------|------------------|-----------------|-------------|---------|
| CPU | FP32 | 0.2191 | 0.3642 | 0.6671 | 0.4468 | 0.5352 |
| GPU | FP32 | 0.2191 | 0.3642 | 0.6671 | 0.4468 | 0.5352 |
| NPU | FP32 | 0.2183 | 0.3669 | 0.6611 | 0.4529 | 0.5375 |
<!-- accuracy-table-end -->

---

## Dig Deeper

Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?

📂 **[View the full project on GitHub](https://github.com/AMD-PAVS/physical_ai_sdk/blob/main/models/easyocr)**

The GitHub repository includes:
- Setup and prerequisites for ROCm environments
- Scripts for CPU, GPU, and NPU runners
- IoU-matching accuracy evaluation pipeline against COCO-Text v2
- Benchmarking and reproduction instructions