Skip to content

High accuracy, explainable, lightweight CNN for terrain recognition.

License

Notifications You must be signed in to change notification settings

deepanshubaghel/Terrain-Recognition

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

99 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Terrain-Recognition

Overview

This project utilizes a Convolutional Neural Network (CNN) to classify terrain types from input images with high precision. The model, trained on 10,000 images across five terrain classes, achieves a validation accuracy of 84.62%, showcasing its strong ability to generalize and accurately distinguish between different terrains.

There is additional a model that achieves 91.93% validation accuracy and 92.40% test accuracy with a size of approximately 15.6 MB, for more details see the EfficientNet-B0 Model README.

Model Architecture

Dataset

We curated a custom dataset by scraping over 10,000 images from the internet across five terrain categories—Coast, Desert, Forest, Glacier, and Mountain—using Python libraries like BeautifulSoup and Selenium. The dataset was meticulously organized into 10,000 training images and 500 test images to ensure diverse coverage for training and robust evaluation.

Sample Data Image

Model Architecture

Total Parameters

  • 6,136,261 (23.41 MB)
  • Trainable params: 6,131,205 (23.39 MB)
  • Non-trainable params: 5,056 (19.75 KB)

Note: The None in the output shape indicates the batch size is flexible and can vary during training or inference.

Model Architecture

Results

Our model achieved an impressive overall accuracy of 84.62% on the test dataset. This success reflects the effectiveness of our carefully designed architecture, which was developed through extensive experimentation and fine-tuning. The high performance underscores the model's strong ability to recognize and classify various terrains.

Model Architecture

Optimization

Through multiple iterations and architectural experimentation, we successfully optimized our model to be significantly more lightweight, reducing its size from 150 MB to just 23 MB. This optimization was crucial for deploying the model in mobile and edge environments where resource efficiency is key.

Model Architecture

This reduction in size makes the model highly suitable for real-world mobile applications without compromising on accuracy, demonstrating the power of a well-optimized architectural design.

Acknowledgments

This project was developed by Team "LearnX" as part of SIH: Deep learning for terrain recognition.

LearnX LEAD : S AKASH

Click to see the list of contributors
  • Vihaan Agrawal

  • Ruchi Chand Thakur

  • Manas Gupta

  • Tanmay Singh

  • Harshith Patnaik

About

High accuracy, explainable, lightweight CNN for terrain recognition.

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 100.0%