Skip to content

Latest commit

 

History

History
149 lines (102 loc) · 5.02 KB

ReadMe.md

File metadata and controls

149 lines (102 loc) · 5.02 KB

Terrain Recognition using Transfer Learning

This project is focused on classifying images into five terrain classes: Coast, Forest, Glacier, Mountain, and Desert. We used transfer learning with ConvNeXT Tiny and MobileNetV2 architectures to leverage pre-trained models for effective terrain classification. This README provides an overview of the models, transformation parameters, accuracy results, and detailed model parameters.

Table of Contents

  1. Project Overview
  2. Dataset and Classes
  3. Model Architecture
  4. Image Preprocessing and Transforms
  5. Model Parameters and Sizes
  6. Results and Evaluation
  7. Conclusion

Project Overview

The primary goal of this project is to perform accurate terrain classification using transfer learning. Two models—ConvNeXT Tiny and MobileNetV2—were chosen for their balance of accuracy and efficiency, and each model was fine-tuned to classify images into five terrain types.

Dataset and Classes

The dataset consists of images belonging to the following classes:

  1. Coast
  2. Forest
  3. Glacier
  4. Mountain
  5. Desert

Each image is preprocessed and transformed to ensure compatibility with the models and maximize classification accuracy.

Model Architecture

1. ConvNeXT Tiny

ConvNeXT Tiny is a convolutional model inspired by the Swin Transformer, optimized for image classification tasks. It provides high accuracy with minimal computational cost.

Model Summary:

  • Architecture: ConvNeXT Tiny
  • Number of Parameters: 27,823,973
  • Input Image Size: 224x224
  • Layer Configuration: Standard ConvNeXT Tiny layers
  • Pre-trained Weights: ImageNet

2. MobileNetV2

MobileNetV2 is a lightweight model designed for efficient use on mobile devices while maintaining high accuracy for tasks like image classification.

Model Summary:

  • Architecture: MobileNetV2
  • Number of Parameters: 2,230,277
  • Input Image Size: 224x224
  • Layer Configuration: Standard MobileNetV2 layers
  • Pre-trained Weights: ImageNet

Image Preprocessing and Transforms

For each model, specific image transformations were applied to prepare the images for the classification task.

ConvNeXT Tiny Transforms

test_transforms = transforms.Compose([
    transforms.Resize((224,224 )),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

MobileNetV2 Transforms

test_transforms = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

Both models use the same normalization parameters, consistent with ImageNet pre-trained weights.

Model Parameters and Sizes

Here is a comparison of the parameters and model sizes:

Model Number of Parameters Model Size (MB)
ConvNeXT Tiny 27.8 million 106 MB
MobileNetV2 2.2 million 9.71 MB
  • Size may vary

Results and Evaluation

ConvNeXT Tiny

  • Accuracy: 92.11%
  • Loss: .2168

Confusion Matrix
Confusion Matrix for ConvNeXT

Model Architecture
ConvNeXT Architecture

Model Accuracy Curve ConvNext Accuracy

MobileNetV2

  • Accuracy: 91.33%
  • Loss: .2798

Confusion Matrix
MobileNetV2 Confusion Matrix

Model Architecture
MobileNetV2 Architecture

Model Accuracy Curve MobileNetV2 Accuracy

Conclusion

This project demonstrates the effective use of transfer learning for terrain classification. While ConvNeXT Tiny offered higher accuracy, MobileNetV2 showed competitive results with a lighter model size, making it suitable for mobile applications. Both models effectively classified the terrains, with each model offering unique advantages depending on the deployment context.

How to Run

  1. Clone the repository.

    git clone https://github.com/Akasxh/terrain-recognition.git
cd terrain-recognition

  2. Run the notebooks

  3. Enjoy the predictions

Feel free to explore, contribute, and leave feedback!