██╗  ██╗██████╗  █████╗ ██╗   ██╗██████╗ ███████╗███████╗███████╗ ██████╗████████╗██╗ ██████╗
╚██╗██╔╝██╔══██╗██╔══██╗╚██╗ ██╔╝██╔══██╗██╔════╝██╔════╝██╔════╝██╔════╝╚══██╔══╝██║██╔═══██╗
 ╚███╔╝ ██████╔╝███████║ ╚████╔╝ ██║  ██║█████╗  █████╗  █████╗  ██║        ██║   ██║██║   ██║
 ██╔██╗ ██╔══██╗██╔══██║  ╚██╔╝  ██║  ██║██╔══╝  ██╔══╝  ██╔══╝  ██║        ██║   ██║██║▄▄ ██║
██╔╝ ██╗██║  ██║██║  ██║   ██║   ██████╔╝███████╗██║     ███████╗╚██████╗   ██║   ██║╚██████╔╝
╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝   ╚═╝   ╚═════╝ ╚══════╝╚═╝     ╚══════╝ ╚═════╝   ╚═╝   ╚═╝ ╚══▀▀═╝

AI-Powered Industrial Radiographic Defect Detection

Manual radiographic inspection is slow, inconsistent, and impossible to scale. XRayDefectIQ changes that.

---

Table of Contents

• What It Does
• System Architecture
• Features
• Model Performance
• Project Structure
• Getting Started
• Tech Stack
• Use Cases
• Roadmap
• License

---

What It Does

XRayDefectIQ is an end-to-end industrial inspection platform that ingests X-ray images, runs them through a trained YOLO detection engine, isolates and characterizes every defect found, and commits structured records to a MySQL database — all through a single interactive UI. No CV expertise required at runtime.

It's built for real NDT workflows: weld inspection, casting analysis, aerospace radiography, and automated manufacturing QA pipelines where accuracy, traceability, and speed are non-negotiable.

---

System Architecture

Text representation

┌─────────────────────────────────────────────────────────────┐
│                      X-ray Image Input                       │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│               Image Enhancement Pipeline                     │
│   Grayscale · CLAHE · Contrast · Sharpen · Denoise · Edge   │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                   YOLO Detection Engine                      │
│          Bounding Boxes · Confidence Scores · Multi-defect  │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│              Defect ROI + Feature Extraction                 │
│   Texture · Entropy · Skewness · Kurtosis · Pixel Stats     │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│                 Inspection Decision Layer                    │
│              Accept / Reject · Inspector Remarks            │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│               MySQL Defect Record Storage                    │
│      Full traceability · Defect ID · Component · Date       │
└─────────────────────────────────────────────────────────────┘

Architecture	Application Flow	Class Structure

---

Features

🔍 Automated Defect Detection

• Upload any X-ray image; YOLO handles the rest
• Annotated output with bounding boxes and per-defect confidence scores
• Multi-defect support — catches everything in a single pass
• ROI extraction per detected defect for downstream analysis

🖼️ Image Processing Suite

Eight enhancement modes to maximize defect visibility before inference:

Mode	Purpose
Grayscale Conversion	Normalize radiographic input
Histogram Equalization	Global contrast correction
CLAHE	Localized adaptive contrast enhancement
Contrast Adjustment	Manual intensity tuning
Sharpening	Edge accentuation for subtle defects
Noise Reduction	Remove sensor and scatter artifacts
Blur Filtering	Smooth out unwanted high-frequency noise
Edge Detection	Highlight structural boundaries

📐 Feature Extraction

For every detected defect region, the platform computes:

• Statistical metrics — Mean, Variance, Skewness, Kurtosis
• Texture descriptors — Entropy, intensity distribution
• Spatial geometry — Bounding box dimensions, centroid, area

🧠 Model Training Console

Full model lifecycle management without leaving the app:

• Upload custom labeled datasets directly from the UI
• Trigger YOLO training jobs and watch live progress
• Save trained checkpoints and hot-swap between models
• Supports iterative retraining as defect libraries expand

🗄️ Inspection Database

Every inspection writes a structured, queryable record:

Field	Description
Defect ID	Unique identifier per detection event
Component Name	Part under inspection
Asset / Satellite	System or assembly context
Component Code	Internal part numbering
Defect Type	Classification label from the model
Accept / Reject	Final disposition
Inspector Remarks	Free-text annotation
Timestamp	Full datetime of inspection

---

Model Performance

Detection metrics from the trained YOLO model evaluated on the held-out test set.

Training Results	Results Overview

Confusion Matrix

Raw	Normalized

Confidence & Recall Curves

F1–Confidence	Precision–Confidence

Recall–Confidence	Precision–Recall

---

Project Structure

XRayDefectIQ/
│
├── app.py                  # Main Streamlit interface — all UI and orchestration logic
├── training_handler.py     # YOLO training pipeline and model management
├── mysql_handler.py        # Database connection, queries, and record insertion
├── DATABASE_SETUP.sql      # Schema creation — run once to initialize
│
├── models/                 # Trained model weights (.pt files)
├── Training images/        # Labeled dataset for custom training runs
└── assets/screenshots/     # Architecture diagrams and performance metrics

---

Getting Started

Prerequisites

• Python 3.9+
• MySQL 8.x (running locally or remotely)

Installation

# 1. Clone the repo
git clone https://github.com/Sanskar121543/XRayDefectIQ.git
cd XRayDefectIQ

# 2. Install dependencies
pip install -r requirements.txt

Configuration

Open mysql_handler.py and update your MySQL credentials before running anything:

host     = "localhost"
user     = "your_user"
password = "your_password"
database = "xraydefectiq"

Database Setup

mysql -u root -p < DATABASE_SETUP.sql

Launch

streamlit run app.py

---

Tech Stack

Layer	Technology
Interface	Streamlit
Detection	Ultralytics YOLO (YOLOv8)
Image Processing	OpenCV, NumPy
Data Handling	Pandas
Storage	MySQL
Language	Python 3.9+

---

Use Cases

• Weld Inspection — Detect porosity, cracks, and incomplete fusion in weld joints
• Casting Analysis — Identify voids, inclusions, and shrinkage defects in metal castings
• Aerospace Radiography — Audit structural components against stringent tolerance specs
• Manufacturing QA — Integrate into production lines for automated pass/fail decisions
• NDT Research — Build and benchmark custom defect detection datasets

---

Roadmap

• Cloud deployment (AWS / GCP / Azure)
• Real-time camera and live feed inspection
• PDF inspection report generation
• Role-based access control (RBAC)
• Explainable AI — Grad-CAM heatmap overlays
• Multi-model ensemble inference

---

Why This Exists

Manual radiographic inspection has three problems: it's slow, it's subjective, and it doesn't scale. A single inspector reviewing hundreds of images per shift introduces fatigue-driven variance. There's no audit trail. There's no feedback loop into process improvement.

XRayDefectIQ is built to fix all three. Fast inference over any X-ray image. Consistent, model-driven detection that doesn't drift with fatigue. And a database-backed record for every inspection, so quality teams can actually learn from the data they're collecting.

---

License

MIT — use it, fork it, build on it.

---

Built by Sanskar Shimpi