---
license: cc-by-sa-4.0
task_categories:
  - image-to-text
language:
  - en
  - bn
  - hi
tags:
  - ocr
  - benchmark
  - assistive-technology
  - multilingual
  - wearable
pretty_name: Assistive OCR Wearable Module — Engine Benchmark Results
size_categories:
  - n<1K
---

# Assistive OCR — Benchmark Results

Real, reproducible benchmark results for the assistive OCR wearable module (offline, multilingual — English, Bengali+English, Hindi+English). This repository holds **results and the ground-truth manifest**, not the source images themselves — the manually-verified image set they were computed from is published separately at [`Arko007/assistive-ocr-data-acquisition`](https://huggingface.co/datasets/Arko007/assistive-ocr-data-acquisition) (folder `manual_verification_99/`), so this repository's numbers can always be traced back to the exact images that produced them.

## What's in this repository

| File | What it is |
|---|---|
| `manual100_final.csv` | The 99-row ground-truth manifest: `id`, `image_path`, `language_profile`, `domain`, `manual_transcript` (each transcript individually re-checked by hand, not auto-generated) |
| `phase4_engine_comparison.json` | Full per-image, per-engine results for an isolated (no cross-engine fallback) comparison of **Tesseract**, **EasyOCR**, and **PaddleOCR** against every row in the manifest |

## Headline results

Isolated per-engine comparison across all 99 manually-verified images (lenient pass = at least one clinically/contextually meaningful word correctly recovered):

| Engine | Attempted | Pass rate | Mean keyword recall | Bengali support |
|---|---|---|---|---|
| Tesseract | 99/99 | 71.7% | 43.2% | Yes |
| EasyOCR | 99/99 | 91.9% | 62.7% | Yes |
| PaddleOCR | 65/99 (34 skipped) | 90.8% (of attempted) | 72.5% (of attempted) | **No** |

**Conclusion:** EasyOCR remains the primary engine and Tesseract the confidence-gated fallback. PaddleOCR cannot process Bengali+English at all (no Bengali in its published model matrix), and even restricted to the languages it does support, it does not meaningfully outperform EasyOCR — so it was evaluated but not adopted. Full reasoning and the fair same-language-subset comparison are in the project's `PROJECT_STATUS.md` (2026-07-31 entry) and `docs/PROJECT_REPORT.md` in the main code repository.

## Reproducing this benchmark yourself on Kaggle

You don't need your own computer to be powerful for this — Kaggle gives you a free notebook with everything already installed. Here's the simplest path, in order.

### Step 1 — Start a new Kaggle notebook

Go to [kaggle.com/code](https://www.kaggle.com/code) → **New Notebook**. In the right-hand sidebar, turn on **Internet** (Settings → Internet → On) so the notebook can install packages and clone the code.

### Step 2 — Add this dataset and the image dataset as notebook inputs

Click **Add Input** (top right) and add:
1. This dataset (`manual100_final.csv`)
2. `Arko007/assistive-ocr-data-acquisition` (the actual images)

They'll appear under `/kaggle/input/` once added.

### Step 3 — Install the OCR engines (one notebook cell)

```bash
!apt-get -qq update && apt-get -qq install -y tesseract-ocr tesseract-ocr-ben tesseract-ocr-hin
!pip install -q "assistive-ocr[tesseract,easyocr] @ git+https://github.com/Bhumika2006-hue/ocr-wearable-module.git"
```

### Step 4 — Point the manifest at the real image paths (one notebook cell)

The manifest's `image_path` column is relative (e.g. `hf_medicines/...`); this cell rewrites it to the actual Kaggle input location so the benchmark can find each image:

```python
import csv, pathlib

IMAGE_ROOT = pathlib.Path("/kaggle/input/assistive-ocr-data-acquisition/manual_verification_99/images")
MANIFEST_IN = "/kaggle/input/<this-dataset-slug>/manual100_final.csv"
MANIFEST_OUT = "/kaggle/working/manifest.csv"

rows = list(csv.DictReader(open(MANIFEST_IN, encoding="utf-8")))
for row in rows:
    row["image_path"] = str(IMAGE_ROOT / pathlib.Path(row["image_path"]).name)

with open(MANIFEST_OUT, "w", newline="", encoding="utf-8") as f:
    writer = csv.DictWriter(f, fieldnames=rows[0].keys())
    writer.writeheader()
    writer.writerows(rows)

print(f"Wrote {len(rows)} rows to {MANIFEST_OUT}")
```

Replace `<this-dataset-slug>` with whatever Kaggle names this dataset once you've added it as an input (visible in the `/kaggle/input/` path shown in the notebook's file browser).

### Step 5 — Run the benchmark (one notebook cell per engine)

```bash
!assistive-ocr-benchmark /kaggle/working/manifest.csv --engine tesseract --output /kaggle/working/tesseract_results.json
!assistive-ocr-benchmark /kaggle/working/manifest.csv --engine easyocr   --output /kaggle/working/easyocr_results.json
```

(Add `--engine paddleocr` too if you also want to reproduce the PaddleOCR comparison — it needs one extra install: `!pip install -q paddlepaddle==2.6.2 paddleocr==2.7.3 "numpy<2"` before running.)

### Step 6 — Look at the results

```python
import json

for name in ("tesseract", "easyocr"):
    data = json.load(open(f"/kaggle/working/{name}_results.json"))
    print(name, "->", data["summary"])
```

Each engine's run produces per-image records plus a `summary` broken down by `language_profile::domain` (word/character error rate, exact-match rate, keyword recall, latency, calibration error). Compare this against the headline table above — it should reproduce those numbers, since both were run from the exact same manifest and images.

### If something doesn't match

- Different Tesseract/EasyOCR versions than the ones the project pinned can shift numbers slightly (a percentage point or two) — this is expected, not a bug.
- If a row errors out instead of producing a number, check that Step 4 actually found the image files (`print(rows[0])` after loading the manifest to sanity-check a resolved path exists with `pathlib.Path(...).exists()`).
- The full setup/troubleshooting guide (for running the whole assistive OCR app, not just this benchmark) is in `docs/SETUP_GUIDE.md` in the main code repository.

## Links

- Main code repository: [Bhumika2006-hue/ocr-wearable-module](https://github.com/Bhumika2006-hue/ocr-wearable-module)
- Source images this benchmark was computed from: [`Arko007/assistive-ocr-data-acquisition`](https://huggingface.co/datasets/Arko007/assistive-ocr-data-acquisition)