Computer Vision + Transfer Learning · Public Kaggle Project

Intel Image Classification

Transfer Learning · EfficientNetB0 · Fine-Tuning

A natural-scene classification project comparing a custom CNN with EfficientNetB0 transfer learning and fine-tuning across six image classes.

RoleDeveloper

Year2025

StatusPublic Kaggle Project

Intel Image Classification transfer learning project cover

OVERVIEW

From a baseline CNN to transfer learning and fine-tuning

This project studies six-class natural-scene recognition using the Intel Image Classification dataset. It compares a CNN trained from scratch with EfficientNetB0 transfer learning and a fine-tuned EfficientNetB0 model.

The workflow includes preprocessing, training, model comparison, confusion-matrix analysis, and visual inspection of correct and incorrect predictions.

SCENE CLASSES

BuildingsForestGlacierMountainSeaStreet

MODEL PROGRESSION

Three progressively stronger training strategies

01Baseline CNN

Custom convolutional network trained from scratch to establish a task-specific baseline.

86.50% Test Accuracy

02EfficientNetB0 Transfer Learning

ImageNet-pretrained EfficientNetB0 used as a feature extractor with an adapted classification head.

92.70% Test Accuracy

03Fine-Tuned EfficientNetB0

Selected pretrained layers are unfrozen and adapted to the scene-classification task.

92.73% Test Accuracy

FINAL RESULT

PROJECT GALLERY

Training evidence and model behavior

Figures below are taken directly from the completed Kaggle notebook and show the dataset, training behavior, model comparison, confusion matrix, and prediction review.

Example images from the six Intel scene classification classes — Dataset · Six Natural Scene Classes

Fine-tuned EfficientNetB0 training and validation curves — Fine-Tuning · Accuracy & Loss

Comparison of baseline CNN, transfer learning, and fine-tuned EfficientNetB0 test accuracy — Model Comparison · Test Accuracy

Confusion matrix for the fine-tuned EfficientNetB0 model — Evaluation · Confusion Matrix

Correct natural scene predictions from the final model — Prediction Review · Correct Examples

Incorrect natural scene predictions used for error inspection — Prediction Review · Incorrect Examples

WHAT THIS PROJECT DEMONSTRATES

Transfer learning as an experimental comparison, not just a final model

The value of the project is the comparison between training from scratch, frozen pretrained features, and controlled fine-tuning. The final notebook also inspects class-level errors rather than relying on accuracy alone.

TECH STACK

PythonTensorFlow/KerasEfficientNetB0Computer Vision

PUBLIC NOTEBOOK

Explore the complete experiment on Kaggle.

The public notebook contains the full preprocessing, training, evaluation, model comparison, and prediction-analysis workflow.

View Kaggle Notebook Back to top