Surface Defect Classifier Training Report

Table of Contents

1. Overview
2. Dataset Description
3. Project Structure
4. Installation & Setup
5. Augmentation and Preprocessing
6. Model and Training Details
7. Performance Metrics Across Epochs
8. Training Metrics Visualization
9. Inference
10. Future Work
11. Summary

---

1. Overview

This project focuses on classification of surface defects in steel manufacturing images as part of a semantic segmentation pipeline. Classification here acts as a preliminary screening mechanism to filter out images with no defects (label 0), thereby reducing computation required during segmentation inference.

We train a ResNet18 model to classify images into 5 classes:

• 0 – No Defect
• 1 to 4 – Corresponding to defect classes in the segmentation dataset

---

2. Dataset Description

The dataset is sourced from a steel surface defect detection competition:

Link: Severstal: Steel Defect Detection on Kaggle

Contents:

• train_images/ – Folder containing training images
• test_images/ – Folder containing inference images
• train.csv – ImageId and defect class mapping

Notes:

• Images not present in train.csv are considered non-defective and assigned label 0
• Multi-class classification is reduced to single-label by selecting the highest defect class per image

---

3. Project Structure

project/
│
├── train.py             # Training entry point
├── inference.py         # Placeholder for inference logic
├── data.py              # Dataset loading & augmentation
├── evaluation.py        # Evaluation functions and metrics
├── models/
│   └── resnet.py        # ResNet18 architecture wrapper
├── utils/
│   └── helpers.py       # Logging, visualization utilities
├── outputs/
│   ├── training\_metrics.png
│   ├── conf\_matrix.png
│   └── roc\_curve.png
├── report.md            # This report
└── requirements.txt     # All dependencies

---

4. Installation & Setup

Prerequisites

• Python 3.8+
• PyTorch >= 1.10
• CUDA GPU recommended for faster training

Installation

git clone https://github.com/iampratyusht/l0-iampratyusht.git
cd l0-iampratyusht
pip install -r requirements.txt

Dataset Preparation

Download the dataset from Kaggle and place folders as:

project/
├── train_images/
├── test_images/
└── train.csv

---

Optional: Manually Creating DataLoaders (For Testing)

Note: You do not need to explicitly create DataLoaders when using train.py — the script handles this internally. This section is only for debugging, testing augmentations, or exploring dataset behavior.

from data import get_dataloaders

train_loader, val_loader = get_dataloaders(
    data_dir="./train_images",
    label_file="./train.csv",
    batch_size=16,
    img_size=(224, 1568),
    num_workers=4
)

This function will:

• Prepare image-level labels (including class 0 for defect-free images)
• Apply augmentations (RandomCrop, Flips, Blackout, etc.)
• Return PyTorch DataLoader objects for training and validation sets

---

5. Augmentation and Preprocessing

During training, we use random crops of size 224x1568. For inference, full-resolution images are used.

Augmentations:

• RandomCrop
• HorizontalFlip, VerticalFlip
• RandomBrightnessContrast (Albumentations)
• Defect Blackout: Known defect pixels are blacked out. If all are removed, label becomes 0.

This simulates natural defects and improves generalization on defect-free images.

---

6. Model and Training Details

We used ResNet18 as the classifier backbone, modified for multi-label classification.

Training Parameters:

• Batch Size: 16 (accumulated to 32)
• Epochs: 10
• Loss Function: BCEWithLogitsLoss
• Optimizer: SGD with momentum

Training Command:

python train.py \
    --model resnet18 \
    --epochs 10 \
    --batch-size 16 \
    --lr 0.01 \
    --data-dir ./train_images \
    --label-file train.csv \
    --save-dir ./outputs

---

7. Performance Metrics Across Epochs

Epoch	Train Loss	Train F1	Train mAP	Train Acc	Train AUC	Val Loss	Val F1	Val mAP	Val Acc	Val AUC
1	0.2770	0.3183	0.4294	0.6190	0.8161	0.2313	0.3339	0.5347	0.7080	0.8973
2	0.2181	0.4052	0.5369	0.7184	0.9005	0.2332	0.4358	0.5695	0.6810	0.8916
3	0.1879	0.4983	0.6315	0.7583	0.9293	0.1787	0.4662	0.6527	0.7876	0.9383
4	0.1667	0.5655	0.6921	0.7843	0.9471	0.1810	0.5887	0.7390	0.7566	0.9472
5	0.1514	0.6492	0.7419	0.8099	0.9567	0.1369	0.6416	0.8630	0.8329	0.9691
6	0.1440	0.6879	0.7732	0.8147	0.9608	0.1888	0.6289	0.8420	0.7677	0.9666
7	0.1377	0.7124	0.7904	0.8288	0.9665	0.1555	0.7275	0.8234	0.8202	0.9630
8	0.1290	0.7233	0.7971	0.8330	0.9695	0.1444	0.6952	0.8238	0.8353	0.9612
9	0.1214	0.7630	0.8316	0.8488	0.9737	0.1826	0.7178	0.8401	0.7979	0.9614
10	0.1098	0.7686	0.8426	0.8608	0.9782	0.1255	0.7934	0.8975	0.8632	0.9793

---

8. Training Metrics Visualization

Training Curve

Displays training and validation loss, F1, mAP, and AUC across epochs.

---

Confusion Matrix

Visualizes true vs. predicted labels for each class.

---

ROC-AUC Curve

Class-wise and macro-average AUC curve visualization.

---

9. Inference

To be filled after inference.py implementation

This section will document:

• Loading the trained model checkpoint
• Preprocessing test images
• Batch-wise prediction with optional TTA (horizontal/vertical flips)
• Saving predicted probabilities

Command:

python inference.py \
    --weights model.pth \
    --image-dir ./test_images \
    --tta hflip,vflip \
    --save-path ./outputs/inference_preds.csv

---

10. Future Work

If time and compute resources allow, we plan to extend this work through:

• Transformer Models for better long-range feature modeling
• Self-Supervised Learning for pretraining on unlabeled industrial images
• Model Ensembling – Combine multiple architectures

---

11. Summary

This project successfully demonstrates a robust ResNet18-based surface defect classifier trained with meaningful augmentations and defect-aware strategies. The classifier improves pipeline efficiency by filtering out defect-absent images and providing high-confidence predictions.

---