🏭 AI Industrial Inspection

Intelligent Surface Defect Detection

Computer Vision · Deep Learning · Transfer Learning · Explainable AI

---

🎯 Project Overview

AI Industrial Inspection is an end-to-end deep learning application built to classify steel surface inspection images into 10 industrial defect categories.

The project uses a trained EfficientNet-B0 model with transfer learning and balanced training. The deployed Streamlit application takes a single uploaded image and returns the predicted defect, confidence score, Top-3 predictions, and a Grad-CAM visual explanation.

Project focus: combining practical image classification, interpretable model output, and a usable web interface into one complete computer-vision workflow.

---

🚀 Live Application

🚀 Launch AI Industrial Inspection

Upload a steel surface image and explore the complete classification + explainability workflow.

---

✨ Key Capabilities

🎯 Multi-Class Classification 10 industrial defect categories The model predicts across the complete supported defect taxonomy.	📈 Confidence Scoring Prediction + confidence The application reports the model confidence associated with the predicted class.
🥇 Top-3 Predictions Ranked alternatives The three highest-probability predictions are displayed for additional context.	🔥 Grad-CAM Explainability Visual model attention The system generates an attention map and overlay for the predicted class.
🖼️ Single-Image Inference End-to-end workflow Upload one steel surface image and receive the complete inspection result.	🌐 Streamlit Deployment Interactive web application The complete inference interface is deployed through Streamlit Community Cloud.

---

📊 Test Performance

87.86% TEST ACCURACY Held-out evaluation

79.61% TEST MACRO F1 10-class performance

83.20% VALIDATION F1 Best validation score

10 DEFECT CLASSES Industrial categories

Evaluation note: These values are the project's existing evaluation metrics. They describe performance on the evaluated dataset and should not be interpreted as guaranteed real-world or production performance.

---

🧩 Supported Defect Classes

01 crease	02 crescent_gap	03 inclusion	04 oil_spot	05 punching_hole
06 rolled_pit	07 silk_spot	08 waist_folding	09 water_spot	10 welding_line

---

⚙️ How the System Works

01 🖼️ IMAGE INPUT Steel surface

→

02 ⚙️ PREPROCESSING Grayscale + resize

→

03 🧠 EFFICIENTNET-B0 Feature extraction

→

04 🎯 PREDICTION Class + confidence

→

05 🔥 GRAD-CAM Visual explanation

---

🏗️ System Architecture

The inference pipeline connects image preprocessing, EfficientNet-B0 classification, ranked prediction output, and Grad-CAM explainability into one inspection workflow.

                 STEEL SURFACE IMAGE
                         │
                         ▼
                ┌─────────────────┐
                │  PREPROCESSING  │
                │ Grayscale+Resize│
                └────────┬────────┘
                         │
                         ▼
                ┌─────────────────┐
                │  EFFICIENTNET-B0│
                │   CLASSIFIER    │
                └────────┬────────┘
                         │
                         ▼
                ┌─────────────────┐
                │   PREDICTION    │
                │ Class+Confidence│
                └────────┬────────┘
                         │
                  ┌──────┴──────┐
                  ▼             ▼
             ┌─────────┐   ┌──────────┐
             │  TOP-3  │   │ GRAD-CAM │
             │ SCORES  │   │ EXPLAIN. │
             └────┬────┘   └─────┬────┘
                  │              │
                  └──────┬───────┘
                         ▼
                EXPLAINABLE RESULT

---

🔍 Explainable AI

From Prediction to Visual Explanation

A predicted label tells us what the classifier selected. Grad-CAM adds a visual explanation by highlighting image regions associated with the model's prediction.

🖼️ Original Inspection input

🔥 Attention Map Grad-CAM output

🌡️ Overlay Attention on image

🎯 Prediction Class + confidence

Interpretation: Grad-CAM is a visual model-explanation technique. It should not be treated as ground-truth defect localization.

---

🖥️ Application Preview

The deployed application brings the complete inspection workflow into a focused dark industrial interface.

01 · Application Dashboard

The landing interface presents the project, model performance, supported classes, and image upload workflow.

---

02 · Inspection Result

The inspection result displays the predicted defect, confidence, model information, inference type, and Top-3 predictions.

---

03 · AI Explainability

The explainability interface presents the original image, Grad-CAM attention map, and AI attention overlay.

---

04 · Supported Defects

The application exposes all 10 supported industrial defect categories.

---

🧠 Model & Methodology

⚡ EfficientNet-B0 Core vision architecture The project uses EfficientNet-B0 as the classification backbone for steel surface defect recognition. Learning strategy Transfer learning adapts the pretrained architecture to the project's 10-class industrial inspection task.

⚖️ Balanced Training Trained inspection model The deployed checkpoint is: best_efficientnet_b0_balanced.pth The Streamlit application loads this trained checkpoint during inference.

---

🛠️ Technology Stack

💻 Core Engineering Layer Technology 🐍 Programming Python 🧠 Deep Learning PyTorch 🏗️ Architecture EfficientNet-B0 🖼️ Image Processing Torchvision · Pillow 📊 Evaluation Scikit-learn

🚀 Application & AI Layer Technology 🔍 Explainability Grad-CAM 📈 Visualization Matplotlib 🌐 Application Streamlit ☁️ Deployment Streamlit Community Cloud 🎯 Inference Single-image classification

Python · PyTorch · EfficientNet-B0 · Torchvision · Grad-CAM · Streamlit

---

🔮 Future Directions

📦 Batch Inspection Extend the current single-image workflow to multiple inspection images.	📍 Defect Localization Explore more precise spatial localization of detected surface defects.	📊 Inspection Analytics Build richer class-level inspection summaries and dashboards.
🗂️ Inspection History Maintain structured records of previous inspection results.	🔔 Review Workflows Add human-review workflows around model predictions.	🏭 Production Integration Explore integration with real industrial inspection pipelines.

These are future development directions and are not represented as currently implemented features.

---

📁 Project Structure

AI_Industrial_Inspection/
│
├── app/
│   └── app.py
│
├── models/
│   └── best_efficientnet_b0_balanced.pth
│
├── notebooks/
│   └── AI_Industrial_Inspection.ipynb
│
├── screenshots/
│   ├── 01_homepage.png
│   ├── 02_inspection_result.png
│   ├── 03_gradcam_explainability.png
│   ├── 04_supported_defects.png
│   ├── 05_system_architecture.png
│   ├── 06_industrial_ai_hero.png
│   ├── 07_explainable_ai.png
│   └── 08_model_technology.png
│
├── src/
│   ├── __init__.py
│   ├── gradcam.py
│   ├── inference.py
│   ├── model.py
│   └── preprocessing.py
│
├── .streamlit/
│   └── config.toml
│
├── .gitignore
├── LICENSE
├── requirements.txt
└── README.md

---

💻 Run Locally

1. Clone the repository

git clone https://github.com/alinkumar/AI_Industrial_Inspection.git
cd AI_Industrial_Inspection

2. Create a virtual environment

Windows PowerShell

python -m venv .venv
.\.venv\Scripts\Activate.ps1

3. Install dependencies

pip install -r requirements.txt

4. Launch the application

streamlit run app/app.py

---

🌐 Live Demo

🚀 AI Industrial Inspection

Explore the deployed application and test the complete inspection workflow.

_{Single Image Inference · Top-3 Predictions · Grad-CAM Explainability}

---

📓 Development Notebook

The complete development notebook is included in:

notebooks/AI_Industrial_Inspection.ipynb

It documents the model development and evaluation workflow used for the project.

---

⚠️ Limitations & Responsible Use

This project is an AI-assisted computer vision portfolio/research implementation, not a certified industrial inspection system.

• Model performance depends on the quality and distribution of inspection images.
• Reported evaluation metrics should not be interpreted as guaranteed production performance.
• Grad-CAM provides model attention visualization, not ground-truth defect localization.
• Real industrial deployment would require domain-specific validation, monitoring, calibration, and operational safety procedures.

The application is designed to demonstrate an end-to-end computer-vision workflow and should not be treated as an autonomous replacement for qualified industrial inspection processes.

---

🔮 Future Directions

📦 Batch Inspection	📍 Defect Localization	📊 Class Dashboards
🗂️ Inspection History	🔔 Review Workflows	🏭 Production Integration

These are future directions and are not represented as currently implemented features.

---

📜 License

This project is released under the MIT License.

See LICENSE for the complete license text.

---

👨‍💻 Author

Alin Kumar

Data Science · Machine Learning · Computer Vision

   

---

🏭 From Surface Image to Explainable AI Decision

EfficientNet-B0 · 10-Class Classification · Confidence Ranking · Grad-CAM · Streamlit

_{Built as a practical Computer Vision & Deep Learning portfolio project.}

⭐ If you found this project useful, consider starring the repository.