Neural Network Image Classification
Building, training and evaluating two different Neural Networks for classifying satellite ground images using TensorFlow
Grade - 1st (96%)
Examples of images from four of the different dataset classes
Accuracy and loss curves for training and validation of CNN
Final accuracy and loss results for training, validation and testing of CNN (significantly better performance than ANN)
Confusion matrix visualisation for CNN testing results (numbers in the diagonal represent correct predictions)
Accuracy and loss curves for training and validation of ANN
Final accuracy and loss results for training, validation and testing of ANN
Confusion matrix visualisation for ANN testing results
Source
This project was made for a Machine Learning course with objective being to build, train and evaluate two different machine learning methods for the task of image classification in Python using TensorFlow and scikit-learn. The images in question being satellite ground images stored as .npy files which can be broken down into 21 different ‘classes‘ with distinct features such as images of suburban areas, coastlines, and airport runways.
I elected to build two different types of Neural Network, a standard Artificial Neural Network (ANN) and a Convolutional Neural Network (CNN). I employed machine learning data augmentation techniques to improve training quality while also introducing methods into each Neural Network architecture to overcome the ‘overfitting’ and ‘underfitting‘ issues that arise when using small image datasets and complex learning algorithms, as was part of the challenge for this project.
I developed the networks over the course of ~2.5 weeks and achieved a final result of 96% for the work (including the report that went with it). Since this isn’t a piece of interactive software to be used by others, it doesn’t really have “interactive features“ per-se, so I’ll summarise the key machine learning takeaways below instead.
Summary
A dataset of 2100 images divided evenly between 21 classes. Images are of 61x61 resolution with 3 colour channels for RGB.
A Ydata label dataset containing class labels for the images.
Data augmentation techniques used to artificially inflate dataset size by 4 times to reduce overfitting and improve train accuracy.
80/10/10 train/test/val split.
22 layer CNN utilising dropout, batch normalisation, kernel regularisation and low learning rate to counter overfitting on small dataset.
4 layer ANN using the Histogram of Oriented Gradients method for feature extraction and Linear Discriminant Analysis for dimension reduction to process image data into an ANN-usable form.