HuaweiOCR

English | 中文

One-line summary

An automated device-label recognition tool: crop fields -> barcode decode -> OCR -> structured export (JSONL), with Windows one-click run.

Overview

This project is a local batch-processing tool for device label / barcode images. It combines object detection (Roboflow inference), barcode decoding, and OCR to extract Model and SN. The core idea is "locate first, then recognize":

• Detect label areas in the full image using a Roboflow model.
• Crop label areas and then crop model/sn sub-fields.
• Run barcode decoding and OCR on the cropped fields.
• Output structured results (JSONL) and debug logs for tracing and tuning.

Pipeline

Input -> crop -> barcode -> ocr -> postprocess -> output

Highlights

• End-to-end batch processing: from raw images to results in one run.
• Multi-stage cropping and filtering to improve accuracy.
• Dual-channel decoding: barcode first, OCR as fallback.
• Debug artifacts saved for easier troubleshooting.
• Local/offline dependencies bundled to reduce external setup.

Core structure

• crop.py: stage 1/2 cropping (Roboflow detect + crop)
• scan2.py: barcode/OCR + structured output
• barcode.py: barcode enhancement pipeline
• run_all.py: one-click pipeline entry
• start.bat: Windows one-click script

Environment

• Windows
• Python 3.10+ (recommended)
• Valid Roboflow API key

Quick start (Windows)

1. Create .env in the project root (do not commit):

API_KEY=your_api_key_here

2. Double-click start.bat.

CLI

python run_all.py --input ./images --out ./out --format jsonl --log-level info --device cpu

Full options:

python run_all.py --help

Pipeline overview

1. Read images from new_images/.
2. Stage 1: detect and crop label areas -> stage1_labels/.
3. Stage 2: crop model/sn fields -> stage2_fields/.
4. scan2.py runs barcode + OCR.
5. Outputs model_sn_ocr.jsonl and debug_ocr_barcode.log.

Outputs

• stage1_labels/: label crops
• stage2_fields/model/: model crops
• stage2_fields/sn/: SN crops
• model_sn_ocr.jsonl: final results (one JSON per line)
• debug_ocr_barcode.log: debug log Default output is repo root; use --out to change it.

Quantitative metrics

CLI prints:

• total images, total time, average time per image
• SN extraction success rate
• regex pass rate
• error distribution (barcode_fail / ocr_fail / regex_fail)

Output format example

Single JSONL line:

{"label_id":"img_001__label_1","model":"S380-S8P2T","sn":"4E25XXXXXXXX","model_src":"barcode","sn_src":"ocr","model_raw":"...","sn_raw":"..."}

Robustness strategies

• Multi-scale upscaling for small barcodes
• ROI cropping to reduce noise
• Rotation attempts (0/90/180/270)
• Regex validation for SN/Model
• Failure samples saved for review

Roadmap

• CSV export and configurable field mapping
• Finer error taxonomy and visual reports
• Incremental CLI processing with resume support
• Lighter releases (optional LFS or model download scripts)

FAQ

• No results: check image clarity, angle, lighting; or tune crop/thresholds.
• API_KEY missing: ensure .env exists and is correct.
• Unstable results: tune parameters in crop.py / scan2.py.

Security

• .env is excluded from the repo, so API keys are not exposed.
• Share .env privately when needed.
• Do not hard-code API keys.
• Key rotation: update API_KEY in .env without code changes.