CNN-Based Image Classification Using Transfer Learning (Materials Science / LPBF SDSS 2507)

This repository shows how to use convolutional neural networks (CNNs) with transfer learning (GoogLeNet) in MATLAB to classify microstructural or microscopy images from additively manufactured (LPBF) AISI 2507 super duplex stainless steel.

The goal is to provide a reproducible, materials-oriented deep learning example: starting from raw microstructure images and ending with a trained model that can tell apart different processing or heat-treatment conditions.

---

1. Background / Motivation

Additively manufactured duplex / super duplex stainless steels exhibit microstructures that vary with:

• laser processing window (power, speed, hatch spacing),
• build orientation,
• and especially post-processing / heat treatment.

These changes appear as visually separable patterns under SEM/optical imaging. A CNN with transfer learning is a good way to:

1. turn these images into labeled data,
2. learn robust features from a pretrained network (GoogLeNet),
3. and test how well the network can discriminate between conditions such as AS, stress-relieved, or solution-annealed samples.

This repo demonstrates exactly that.

---

2. What’s Inside

• README.md — this file.
• Transfer Learning(GoogLeNet) for Microstructure Classification of Additively Manufactured AISI 2507 SDSS Samples — MATLAB script/notebook illustrating the full transfer-learning workflow (load images → split → augment → retrain final layers → evaluate).
• LICENSE — MIT, so you can reuse and extend.
• (Optional / to be added) /dataset/ or similar folder for image data grouped by class.

Note: GitHub currently shows the project as “Digital Image Processing & Deep Learning for Material Science,” so you can extend this repo later with other course projects. :contentReference[oaicite:2]{index=2}

---

3. Requirements

• MATLAB (R2022a or newer recommended)
• Deep Learning Toolbox
• Image Processing Toolbox (for reading/augmenting microscopy images)
• Computer Vision Toolbox (optional but useful)
• A GPU will speed up training, but CPU works for small datasets.

---

4. Dataset Expectations

This project assumes your images are organized in subfolders by class, e.g.

data/
├── AS/
│   ├── img_001.png
│   ├── img_002.png
├── SR400_1h/
│   ├── img_010.png
├── SR500_1h/
│   ├── ...
└── SA1100_15min/
    ├── ...