---

📋 Table of Contents

• Overview
• Key Results
• Architecture
• Features
• Tech Stack
• Installation
• Usage
• Pipeline Details
• File Structure
• Dataset & Databases
• Contributors
• License

---

🎯 Overview

An OCR-based pipeline built for IISc's CGPL (Combustion, Gasification & Propulsion Laboratory) to automate combustion gas composition analysis from mass spectrometry outputs. The system processes PDF or image inputs containing mass spectra, extracts m/z peak labels using computer vision, and identifies the closest matching chemical compound from curated gas databases — returning confidence scores per identified species.

The pipeline handles two fundamentally different spectra types:

• Clean integer references — standard reference spectra with integer m/z values
• Noisy instrument decimals — raw instrument output with values like 27.9012 requiring normalization

This distinction drove the core engineering challenge: designing a rounding and matching strategy robust enough to bridge both formats reliably, reducing manual peak identification effort for active gasification experiments.

---

📊 Key Results

Metric	Value
m/z peak extraction accuracy	85%
Spectra types handled	2 (integer reference + noisy instrument)
Databases matched against	2 (pure elements + 25 combustion samples)
Output format	Percentage confidence scores per identified gas

• Input formats: PDF spectra files and PNG/image inputs
• Database coverage: Pure elemental gases + 25 combustion compound samples

---

🧠 Architecture

Mass Spectrometry Output (PDF / Image)
          │
          ▼
  ┌───────────────────┐
  │  PDF → Image      │  Convert spectra pages to rasterised images
  │  Preprocessing    │  for downstream vision processing
  └───────────────────┘
          │
          ▼
  ┌───────────────────┐
  │  Hough-Transform  │  Detects X and Y axes from spectrum graph
  │  Axis Detection   │  via OpenCV — adaptive graph cropping
  │  (OpenCV)         │  isolates the m/z peak region
  └───────────────────┘
          │
          ▼
  ┌───────────────────┐
  │  EasyOCR          │  Reads m/z peak labels from the cropped
  │  Peak Extraction  │  spectrogram region
  └───────────────────┘
          │
          ▼
  ┌───────────────────┐
  │  Decimal          │  Custom rounding logic (≥ 0.6 rounds up)
  │  Normalisation    │  normalises instrument noise — bridges
  │                   │  integer references and raw decimals
  └───────────────────┘
          │
     ┌────┴────────────────┐
     ▼                     ▼
  Pure Elements DB    Combustion Samples DB
  (reference gases)   (25 compounds)
     │                     │
     └────────┬────────────┘
              ▼
  ┌───────────────────┐
  │  Confidence Score │  Percentage match score returned
  │  Output           │  per identified gas species
  └───────────────────┘

---

✨ Features

• Dual spectra handling — processes both clean integer references and noisy instrument decimal spectra without configuration changes
• Hough-transform axis detection — automatically detects graph axes for adaptive cropping, requiring no manual region annotation
• EasyOCR peak extraction — reads m/z values directly from spectrum images, tolerating real-world print quality variation
• Custom decimal normalisation — ≥ 0.6 rounds up logic bridges instrument noise to integer references
• Two-database matching — matches against pure elements and a curated 25-compound combustion sample library
• Confidence scores — returns percentage confidence per identified gas species for downstream analysis
• PDF and image input support — accepts raw instrument PDF exports and pre-rasterised images

---

🛠️ Tech Stack

Layer	Tools
OCR	EasyOCR
Computer Vision	OpenCV · Hough Transform
Image Processing	PIL / Pillow · NumPy
Data Handling	Pandas · openpyxl
Input Formats	PDF (via pdf2image / PyMuPDF) · PNG
Language	Python

---

🔧 Installation

git clone https://github.com/ShettyShravya03/chemical-compound-recognition.git
cd chemical-compound-recognition

pip install -r requirements.txt

Key dependencies: easyocr opencv-python numpy pandas Pillow pdf2image openpyxl

Note: EasyOCR downloads model weights on first run (~200 MB). Ensure an internet connection on initial setup.

---

🚀 Usage

Run on a PDF spectrum file

python easyocr_match.py --input Spectra\ 1.pdf

Run on an image input

python simple.py --input page_1.png

Compare across all samples

python final_comparison.py

Sample output:

Extracted m/z values: [28, 32, 44, 14, 16]
Best match: Carbon Dioxide (CO₂)
Confidence: 87.3%

Runner-up: Nitrogen (N₂)
Confidence: 61.4%

Match against a specific database

from easyocr_match import match_spectrum

result = match_spectrum(
    input_path="Spectra 3.pdf",
    database="combustion"   # or "elements"
)

print(result["compound"])     # e.g. "Carbon Monoxide"
print(result["confidence"])   # e.g. 0.83
print(result["matched_peaks"])

---

🔍 Pipeline Details

Stage 1 — PDF / Image Ingestion

Input spectra arrive as PDF exports from mass spectrometry instruments or as pre-rasterised images. PDFs are converted page-by-page to high-resolution images for downstream processing.

Stage 2 — Axis Detection (Hough Transform)

OpenCV's Hough Line Transform detects the X (m/z) and Y (intensity) axes of the spectrum graph. Adaptive graph cropping then isolates only the peak label region, reducing OCR noise from surrounding annotations, titles, and legends.

Stage 3 — OCR Peak Extraction (EasyOCR)

EasyOCR reads the m/z peak values from the cropped region. This stage handles two spectra types:

• Integer references: Standard clean outputs — 28, 32, 44
• Instrument decimals: Raw noisy readings — 27.9012, 31.9867, 43.9901

Stage 4 — Decimal Normalisation

A custom rounding rule (decimal part ≥ 0.6 → round up, otherwise round down) normalises instrument noise to integer m/z values, enabling matching against integer-indexed reference databases without manual correction.

Stage 5 — Database Matching

Normalised m/z sets are matched against:

Database	Contents
Pure elements	Reference spectra for elemental gases
Combustion samples (SAMPLES.xlsx)	25 combustion compound spectra

Match confidence is computed as the percentage of extracted peaks present in the candidate compound's reference spectrum — the highest-scoring compound is returned as the primary identification.

Scripts Reference

Script	Purpose
easyocr_match.py	Main OCR extraction + database matching
simple.py	Lightweight single-image extraction
match_simple.py	Simplified matching against element DB
final_comparison.py	Batch comparison across all sample spectra

---

📁 File Structure

.
├── Images/                # Reference spectrum images
├── Samples-25/            # 25 combustion compound sample spectra
├── Spectra 1.pdf          # Example instrument spectra (1–5)
├── Spectra 2.pdf
├── Spectra 3.pdf
├── Spectra 4.pdf
├── Spectra 5.pdf
├── Carbondioxide.pdf      # Reference spectra for pure elements
├── Nitrogen.pdf
├── Oxygen.pdf
├── SAMPLES.xlsx           # Combustion samples database
├── ONSpectra.xlsx         # O₂ / N₂ reference spectra
├── Book.xlsx              # Supplementary data
├── page_1.png             # Example rasterised spectrum page
├── easyocr_match.py       # Main OCR + matching pipeline
├── final_comparison.py    # Batch comparison script
├── match_simple.py        # Simplified matching utility
├── simple.py              # Lightweight image extraction
├── requirements.txt
└── README.md

---

📂 Dataset & Databases

Instrument Spectra

Five real mass spectrometry outputs (Spectra 1–5.pdf) from IISc's CGPL gasification experiments, covering combustion gas mixtures including CO₂, CO, N₂, O₂, and hydrocarbons.

Reference Databases

Database	Format	Compounds
Pure elements	Individual PDFs (N₂, O₂, CO₂) + ONSpectra.xlsx	Elemental gases
Combustion samples	SAMPLES.xlsx	25 combustion compounds

Note: The sample spectra in this repository are from real gasification experiments at IISc CGPL. Full datasets remain under the lab's data governance policy.

---

👥 Contributors

Built during a Software Engineering Internship at IISc's CGPL Lab, Bengaluru (Jun–Jul 2025).

Contributor	GitHub
Shravya S Shetty	@ShettyShravya03
Sridevi	@Sridevi25git
Khushi R Shetty	@KhushiRShetty

---

📜 License

MIT © 2025 Shravya S Shetty, Sridevi, Khushi R Shetty

_{Built at IISc CGPL Lab, Bengaluru · Software Engineering Internship 2025}